Performance Evaluation of YOLO11 and YOLO26 for Detection of Low-Contrast Surface Contamination on Eggshell using Fluorescence Imaging

Performance Evaluation of YOLO11 and YOLO26 for Detection of Low-Contrast Surface Contamination on Eggshell using Fluorescence Imaging | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Performance Evaluation of YOLO11 and YOLO26 for Detection of Low-Contrast Surface Contamination on Eggshell using Fluorescence Imaging Alin Khaliduzzaman, Tanea Gray, Mamunur Rahman, Naoshi Kondo, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9362909/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 4 You are reading this latest preprint version Abstract Ensuring eggshell surface cleanliness is important for product quality and food safety. Routine inspection under visible illumination can identify gross residues but may miss thin or low-contrast surface films that are difficult to perceive consistently, particularly at industrial speeds. Eggshell surfaces can also carry microorganisms, motivating improved, non-destructive sensing methods. In this study, we evaluated fluorescence imaging as a contrast-enhancement modality for visualizing fecal-smear contamination not readily apparent under standard illumination, and we developed a computer vision workflow for automated detection. Fluorescence images were annotated in CVAT and used to train Ultralytics YOLO11s and YOLO26s object detection models implemented in PyTorch. Images were partitioned into training (70%) and validation (30%) sets. The models’ performance was assessed on a validation set using precision, recall, and mean average precision (mAP50 and mAP50-95). On the validation set, the YOLO11 model achieved little higher accuracy (mAP50 = 0.995 and mAP50-95 = 0.995) but inference time was higher compared to YOLO26 for distinguishing clean shells from fecal-smear contamination under the study’s imaging conditions. Inference time was 194.6ms for YOLO11 whereas 164.3ms for YOLO26 per image on a CPU (Intel i7-12700). These results indicate that fluorescence-assisted imaging combined with deep-learning can support rapid, non-destructive screening for eggshell surface residues and merits further validation using larger datasets that vary by contamination presentation, imaging conditions, and independent production lots. Eggshell surface quality computer vision object detection egg industry 4.0 Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Introduction Ensuring eggshell surface cleanliness is very important for egg and egg products quality compliance as it harbors microorganisms which can further penetrate inside and cause serious food safety issues (Aygun, 2017; De Reu et al., 2006). It can also cause increased mortality of embryos, lower hatchability, and increased early chick mortality (J. Svobodová, 2014). Routine inspection under visible illumination can identify gross residues but may miss thin or low-contrast surface films that are difficult to perceive consistently, particularly at industrial speeds which motivating improved, non-destructive sensing methods. In this study, we evaluated fluoresceine imaging as a contrast-enhancement modality for visualizing fecal-smear contamination not readily apparent under standard illumination, and we developed a computer vision workflow for automated detection. Egg industry is currently undergoing a significant transformation with the adoption of Industry 4.0 technologies, referred to as Egg Industry 4.0 (EI 4.0), which emphasizes automation and digitalization to enhance efficiency and product quality (Ahmed et al., 2023). The main concept of industry 4.0 is digitization of industrial automation system to ensure quality of foods rather than quantity whereas industry 5.0 is mostly focused on sophisticated or highly precise grading considering aesthetic value of food. Traditionally, eggs have been checked visually by product line inspector for dirt and defects, but this approach is slow and prone to error, especially when large numbers of eggs need to be inspected. Some egg industry applied computer vision based visible defect or size and shape-based grading. But some of the dirt on shells remains invisible by naked eye which could compromise the consumer or market standards. Such surface contamination or foreign matters on eggs surface could be visible outside human vision range under UV light and using appropriate optics. Recently, fluorescence techniques, especially imaging, are getting attention of researchers for potential application in agricultural sciences for quality evaluation (Khaliduzzaman et al., 2023; Konagaya, Omwange, et al., 2020). The techniques involve the detection of naturally occurring fluorescent compounds known as fluorophores which emit light of longer wavelengths upon absorbing energy of shorter wavelengths. FL-imaging provides spatial information on how the fluorophores are distributed and localized in the target sample. This imaging technique provides more specific information than FL-spectroscopic techniques. The most common imaging technique that is used in agricultural sciences is ultraviolet-induced visible fluorescence imaging having excitation wavelength in the UV region and emission in the visible wavelength range (Konagaya, Al Riza, et al., 2020). Thus, the fluorescence imaging of egg surface could visualize this invisible dirt under UV lighting and can be further detected in real time by deploying deep learning object detection like FR-CNN and YOLO. YOLO11 is the latest iteration in the Ultralytics YOLO series for real-time object detection which has proven superior performance over other detectors (Jocher et al., 2023). YOLO11 introduces significant improvements in architecture and training methods, making it a versatile choice for a wide range of computer vision tasks. These improvements include enhanced features extraction by applying improved backbone and neck architecture, optimized for efficiency and speed up, greater accuracy with lower number of parameters, adaptability across environment for seamlessly deployment and broad range of supported tasks (e.g., instance segmentation, image classification, pose estimation or oriented object detection). Therefore, the aim of this research is to classify clean and dirt eggs non-destructively for industrial grading using UV imaging combined with YOLO11. On the other hand, YOLO26, released in late 2025/early 2026, focuses on extreme deployment efficiency by shifting to an end-to-end, NMS-free design, eliminating the post-processing bottlenecks common in YOLO11 (Jocher & Qiu, 2026). Therefore, fluorescence-assisted imaging combined with deep-learning can support rapid, non-destructive screening for eggshell surface residues and merits further validation using larger datasets that vary by contamination presentation, imaging conditions, and independent production lots. Materials and Methods This experiment was conducted in accordance with requirements of ethics committee of the Kyoto University, Japan (Approval number: 28–59). Imaging system and sample acquisition The images of light brown eggs were taken in a dark room using EOS DSLR (kiss x7, Color Inc., Japan) under white and UV light (365 nm LEDs) for color and fluorescence imaging. The EOS DSLR camera was placed 200 mm above the sample. All images were taken with same F-number (f/6.3) and ISO (100). The sample was illuminated using two white bar LEDs (LDL2-80 × 16SW2, CCS Inc. Japan) with a shutter speed of 1/5 s for normal color images and two 365 nm UV bar LEDs (LDL-71 × 12UV2-365-N, CCS Inc. Japan) with a shutter speed of 2.5 s for fluorescence images of eggs. The distance between the sample and the lighting devices was set as 160 mm (white bar LEDs), and 180 mm (365 nm bar LEDs) for the UV-imaging system. The typical UV images of clean and contaminated eggs are shown in Figure 1. Dataset preparation and annotation Fluorescence images ( n = 74; one image per egg) were annotated in Computer Vision Annotation Tool (CVAT) for dataset preparation. The dataset was split into a 70:30 ratio for training and validation, respectively. The images used for validation were not included in the training set. Two classes (clean and surface contaminated) were defined as Clean_Egg and Dirty_Egg. YOLO detectors and training configuration The YOLO11s and YOLO26s models were implemented utilizing the PyTorch framework on a personal computer (12th Gen Intel(R) Core-i7, 2.10 GHz, 64-bit). Some key hyperparameters used during training and inference are summarized in Table 1. Table 1. Hyperparameter configuration for YOLO11s and YOLO26s in eggshell contamination detection. Parameter Value Description Image Size 640 X 640 Input resolution for training and inference. Batch Size 16 Number of samples processed before the model is updated. Epochs 100 Total number of complete passes through the training dataset. Initial Learning Rate (lr 0 ) 0.01 Step size at the beginning of training. Momentum 0.937 Factor to accelerate gradients in the right direction. Weight Decay 0.0005 Regularization term to prevent overfitting. Optimizer Auto Auto (SGD for YOLO11; MuSGD for YOLO26) Box Loss Gain 7.5 Weighting factor for bounding box regression loss. IoU Threshold 0.7 Intersection over Union threshold for training anchor assignment. YOLO26 removes Distribution Focal Loss (DFL), introduces a redesigned backbone, and uses the MuSGD optimizer to achieve up to 43% faster CPU inference and improved small-object detection. Some basic differences between YOLO11 and YOLO26 are shown in Figure 2. YOLO-based inference workflow Figure 3 presents the YOLO-based inference pipeline adopted in this study. Images are resized and processed through a convolutional backbone to extract spatial features, which are used to predict object locations and class labels. Final detections are obtained via post-processing, using NMS in YOLO11 and an end-to-end prediction strategy in YOLO26. Performance metrics Model performance was evaluated using precision (P), recall (R), average precision (AP), mean average precision at IoU = 0.50 (mAP50) and mean average precision averaged over IoU thresholds from 0.50 to 0.95 (mAP50–95) (Khaliduzzaman et al., 2021; Rahman et al., 2025). Precision is used to measure how well the predictions of the bounding box are correct with the dataset. It is computed by the ratio of actual positive predictions as shown in Eq. 1. where, TP and FP are known as true positive, false positive values respectively. Recall represents the potential of the model to have the correct prediction of the bounding box measurements within the dataset. It is computed as a ratio of true positive predictions (Eq. 2). where, TP and FN are known as true positive, false negative values respectively. Average Precision (AP) summarizes the precision-recall curve for a single class. It is calculated as the area under the precision-recall curve (Eq. 3), integrating precision P(R) as a function of recall, R: Mean Average Precision (mAP) assesses the overall detection accuracy across all classes. mAP50 is the average precision calculated at a specific Intersection over Union (IoU) threshold of 0.50. It indicates the model's consistency in detecting objects with moderate overlap. mAP50-95 is the average of mAP values computed at IoU thresholds from 0.50 to 0.95 in steps of 0.05. High mAP50-95 reflects the model's robustness in precise object localization. The formula for mAP is shown in Eq. 4: Results and Discussion Although YOLO11s model performed better in terms of accuracy (i.e., precision, recall and thus mAP), YOLO26s model had faster inference time as shown in Table 2. On the validation set, the YOLO11 model achieved higher accuracy of mAP50 = 0.995 and mAP50-95 = 0.995 (precision = 0.908; recall = 0.996) for distinguishing clean shells from fecal-smear contamination whereas YOLO26s model on the same validation set achieved mAP50 = 0.966 and mAP50-95 = 0.966 with precision = 0.834 and recall = 0.946. It was found that YOLO11s was able to detect darts eggs as dart eggs more accurately and it was opposite for the YOLO26s model (detected clean eggs as clean eggs). Table 2 Model evaluation metrics for YOLO11s and YOLO26s validation set Model Precision Recall mAP50 mAP50-95 Inference Time (ms) YOLO11s 0.908 0.996 0.995 0.995 194.6 YOLO26s 0.834 0.946 0.966 0.966 165.3 Inference time was 194.6ms for YOLO11s whereas 164.3ms for YOLO26s per image on a CPU (Intel i7-12700). The reasons behind the faster inference in CPU are that YOLO26 used Ultralytics edge-optimized model which eliminated NMS with a native end-to-end predictor; removed DFL (Distribution Focal Loss) for simpler, faster inference; introduced MuSGD optimizer (SGD+Muon hybrid) for stable and quick convergence results in faster CPU inference for deployment on low-power devices (Sapkota et al., 2026). These deployment-first design choices align with broader trends in food/agriculture machine vision where inference efficiency and robustness are critical for real-time inspection and sorting pipelines (Shen et al., 2024). Most of the predictions are found to be on the diagonal as indicated by the confusion matrices (Figure 4) implying that the models tend to predict the correct class when a detection is being made. The off-diagonal values indicate the most common practical failure modes, contaminations that are predicted to be clean (false negatives), and clean shells that are predicted to be contaminated (false positives). The off-diagonal counts in YOLO11s are lower than those in YOLO26s, which corresponds to the fact that its precision and recall was higher (Table 2). False negatives tend to be the more expensive error in an inspection process since an egg with contamination can be missed; the increased recall of YOLO11s justifies harsher screening conditions by the conditions of the study. One of the potential reasons for misclassification was that few egg samples contained few shiny dots or white dots (as a contamination) looked very close to clean egg. The loss curves (Figure 5) give additional reasons as to why YOLO26s is efficient and can be optimized. Throughout training, YOLO26s has smoother and more steady reduction of training and validation losses whereas YOLO11s has stronger spikes in validation losses. Such spikes are common in small or moderately sized datasets, and this suggests that the model is sensitive to a limited number of hard/ambiguous examples (e.g., very thin smears, uneven fluorescence intensity, specular highlights or borderline labeling cases). Notably, YOLO26 is specifically crafted with training-stability techniques, such as ProgLoss and STAL, which are intended to stabilize learning and enhance small/weak-target assignment, as well as MuSGD for better convergence. In a nutshell, while YOLO11 is a capable, hybrid-task model, YOLO26 is fundamentally redesigned as an edge-first model, focusing on removing bottlenecks for faster, more efficient, and more accurate deployment. One misclassification was due to very few shiny dots or white dots (as contamination) on the eggshell surface which looked very close to clean egg. Hence, egg grader company can consider an omittable threshold to define a clean egg and a surface contaminated egg. Therefore, YOLO26 could be a good choice for real time detection of invisible or low contrast surface contamination of egg at industrial settings. Conclusions Eggshell surface contamination can be detected in real-time scenario using fluorescence imaging combined with YOLO11 or YOLO26, as both the models have very good performance on detection of clean and low contrast surface contamination. However, the YOLO26 model might be good for egg industry 4.0 considering faster inference time which is a very important factor for industrial grading application. On the validation set, the YOLO11 model achieved slightly higher accuracy (i.e., mAP50 = 0.995 and mAP50-95 = 0.995) than YOLO26 model for distinguishing clean shells from fecal-smear contamination. On the other hand, YOLO26 performed faster as inference time was 194.6ms for YOLO11 while 164.3ms for YOLO26 per image on a CPU (Intel i7-12700) setting. These results indicated that fluorescence-assisted imaging combined with deep-learning can support rapid, non-destructive screening for eggshell surface residues and merits further validation using larger datasets that vary by contamination presentation, imaging conditions, and independent production lots. Declarations Funding Declaration Financial support was provided by Grant-in-Aid for JSPS Postdoctoral Research Fellow under the JSPS KAKENHI Grant Number 19F19396 and Project Number 82019000018. Ethics Declaration Although research manuscripts didn’t use any data of any live animal, this experiment was conducted in accordance with requirements of ethics committee of the Kyoto University, Japan (Approval number: 28–59). Ethics committee (animal research) in Kyoto University follows internationally recognized guidelines for animal experiments. Competing interests Alin Khaliduzzaman is on the editorial board member of this journal. Author contributions Alin Khaliduzzaman: Conceptualization, Methodology, Investigation, Data curation, Formal analysis, Writing – original draft. Tanea Gray: Formal analysis, Mamunur Rahman: Writing-review and editing. Naoshi Kondo: Data-curation, Project administration. Isabella Condotta: Writing – review and editing, Research funding. References Ahmed, M. W., Hossainy, S. J., Khaliduzzaman, A., Emmert, J. L., & Kamruzzaman, M. (2023). Non-destructive optical sensing technologies for advancing the egg industry toward Industry 4.0: A review. In Comprehensive Reviews in Food Science and Food Safety (Vol. 22, Number 6, pp. 4378–4403). John Wiley and Sons Inc. https://doi.org/10.1111/1541-4337.13227 Aygun, A. (2017). Chapter 13 - The Eggshell Microbial Activity. In P. Y. Hester (Ed.), Egg Innovations and Strategies for Improvements (pp. 135–144). Academic Press. https://doi.org/https://doi.org/10.1016/B978-0-12-800879-9.00013-5 De Reu, K., Grijspeerdt, K., Messens, W., Heyndrickx, M., Uyttendaele, M., Debevere, J., & Herman, L. (2006). Eggshell factors influencing eggshell penetration and whole egg contamination by different bacteria, including Salmonella enteritidis. International Journal of Food Microbiology , 112 (3), 253–260. https://doi.org/https://doi.org/10.1016/j.ijfoodmicro.2006.04.011 J. Svobodová, E. T. (2014). FACTORS AFFECTING MICROBIAL CONTAMINATION OF MARKET EGGS: A REVIEW. Scientia Agriculturae Bohemica , 45 (4), 226–237. Jocher, G., Chaurasia, A., & Qiu, J. (2023). Ultralytics YOLOv8 . https://github.com/ultralytics/ultralytics Jocher, G., & Qiu, J. (2026). Ultralytics YOLO26 . https://github.com/ultralytics/ultralytics Khaliduzzaman, A., Fujitani, S., Kashimori, A., Suzuki, T., Ogawa, Y., & Kondo, N. (2021). Non-invasive Detection of Chick Embryo Gender Based on Body Motility and a Near-infrared Sensor. In EAEF (Vol. 14, Number 2). Khaliduzzaman, A., Omwange, K. A., Al Riza, D. F., Konagaya, K., Kamruzzaman, M., Alom, M. S., Gao, T., Saito, Y., & Kondo, N. (2023). Antioxidant assessment of agricultural produce using fluorescence techniques: a review. In Critical Reviews in Food Science and Nutrition (Vol. 63, Number 19, pp. 3704–3715). Taylor and Francis Ltd. https://doi.org/10.1080/10408398.2021.1992747 Konagaya, K., Al Riza, D. F., Nie, S., Yoneda, M., Hirata, T., Takahashi, N., Kuramoto, M., Ogawa, Y., Suzuki, T., & Kondo, N. (2020). Monitoring mature tomato (red stage) quality during storage using ultraviolet-induced visible fluorescence image. Postharvest Biology and Technology , 160 . https://doi.org/10.1016/j.postharvbio.2019.111031 Konagaya, K., Omwange, K. A., Al Riza, D. F., Khaliduzzaman, A., Martínez Oliver, A., Rovira-Más, F., Nagasato, H., Ninomiya, K., & Kondo, N. (2020). Association of fruit, pericarp, and epidermis traits with surface autofluorescence in green peppers. Photochemical and Photobiological Sciences , 19 (12), 1630–1635. https://doi.org/10.1039/d0pp00236d Rahman, M., Souza, V. H. S., Brown-Brandl, T. M., Rohrer, G. A., Shi, Y., & Condotta, I. C. F. S. (2025). 56. Automated monitoring of sow nursing behaviors and postures in farrowing crates through computer vision techniques. Animal - Science Proceedings , 16 (4), 595–596. https://doi.org/https://doi.org/10.1016/j.anscip.2025.08.211 Sapkota, R., Cheppally, R. H., Sharda, A., & Karkee, M. (2026). YOLO26: Key Architectural Enhancements and Performance Benchmarking for Real-Time Object Detection . http://arxiv.org/abs/2509.25164 Shen, C., Wang, R., Nawazish, H., Wang, B., Cai, K., & Xu, B. (2024). Machine vision combined with deep learning–based approaches for food authentication: An integrative review and new insights. In Comprehensive Reviews in Food Science and Food Safety (Vol. 23, Number 6). John Wiley and Sons Inc. https://doi.org/10.1111/1541-4337.70054 Additional Declarations No competing interests reported. Cite Share Download PDF Status: Under Review Version 1 posted Reviewers invited by journal 24 Apr, 2026 Editor assigned by journal 10 Apr, 2026 Submission checks completed at journal 10 Apr, 2026 First submitted to journal 08 Apr, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9362909","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":634014384,"identity":"8a92b28e-2ebe-409e-a46d-36faf3424682","order_by":0,"name":"Alin Khaliduzzaman","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABT0lEQVRIie2PP0vDQBiHrxRuOu164ST3CYS3FLq0pN/C+Q2BZEmgIBQnqQh2EboGMvgVUgJ1PQh0sTgXuqSLk4VAl4JCTQKCrVF0E8kzHPf+efjxElJR8QfhnNQU0veyW7y1BIHq+U8lpQr5oNgkq0kd0v5xq1DwC4UcKFTzU90c5o0SRQuuE5UM4p4cPZibPnYv5WgOgkHbuTuLV1mKoZ8O9xRxMgOFj3G9dutFwkebw9yFFgPbmyztbESsVlvtKTpHUOZNTOvEC8XRNuZAXLAYzLxJgLmizOmh4qS5wmhjHb0w3HE5foaYwc5pBtmoRBHcLVI4495UMFScLNzmlQ8UpXBLUzTf7We3OMD5etphaGnh4umcpECbochH8OkWvnAmq+2g05NjL1oyNBpybN1v8ZVKGThRml4Y+oHyDVBswk/Xc+TwN9sVFRUV/5g38X5/hcnCkRQAAAAASUVORK5CYII=","orcid":"","institution":"University of Illinois Urbana-Champaign","correspondingAuthor":true,"prefix":"","firstName":"Alin","middleName":"","lastName":"Khaliduzzaman","suffix":""},{"id":634014389,"identity":"513d1961-f519-427b-be4e-255a0499b1f4","order_by":1,"name":"Tanea Gray","email":"","orcid":"","institution":"Langston University","correspondingAuthor":false,"prefix":"","firstName":"Tanea","middleName":"","lastName":"Gray","suffix":""},{"id":634014390,"identity":"f8dbcb10-0c68-4f1a-a4b0-e3fb9214f15a","order_by":2,"name":"Mamunur Rahman","email":"","orcid":"","institution":"University of Illinois Urbana-Champaign","correspondingAuthor":false,"prefix":"","firstName":"Mamunur","middleName":"","lastName":"Rahman","suffix":""},{"id":634014393,"identity":"d5033ffc-aa8c-440a-bc90-d8b4ad6e7964","order_by":3,"name":"Naoshi Kondo","email":"","orcid":"","institution":"Kyoto University","correspondingAuthor":false,"prefix":"","firstName":"Naoshi","middleName":"","lastName":"Kondo","suffix":""},{"id":634014394,"identity":"37461f4f-bab1-4ce7-88fd-08518e518266","order_by":4,"name":"Isabella Condotta","email":"","orcid":"","institution":"University of Illinois Urbana-Champaign","correspondingAuthor":false,"prefix":"","firstName":"Isabella","middleName":"","lastName":"Condotta","suffix":""}],"badges":[],"createdAt":"2026-04-09 04:09:27","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9362909/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9362909/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":108837044,"identity":"db0b4cdc-cce6-4c93-994c-6aa4eeece61d","added_by":"auto","created_at":"2026-05-09 00:04:30","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":345138,"visible":true,"origin":"","legend":"\u003cp\u003eTypical fluorescence images of (a) clean (ID: IMG_0588) and (b) contaminated eggs (ID: IMG_0484) c) RGB image with very low contrast contamination and d) FL image of the same egg with visible surface contamination.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-9362909/v1/c2adeab1e9be304a2ff6e6d7.png"},{"id":108837041,"identity":"13394603-6163-40c0-940c-a58ed91e652c","added_by":"auto","created_at":"2026-05-09 00:04:29","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":137173,"visible":true,"origin":"","legend":"\u003cp\u003eFundamental differences between YOLO11 and YOLO26\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-9362909/v1/767fd8290c4b7ffdf8923eeb.png"},{"id":108976986,"identity":"a4cdb38a-62b6-4267-ba25-cc122b1c5acf","added_by":"auto","created_at":"2026-05-11 11:29:49","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":265831,"visible":true,"origin":"","legend":"\u003cp\u003eYOLO-based general image processing flow for detection of surface contaminated and clean eggs.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-9362909/v1/c504a03b465fad400cf9aead.png"},{"id":108837045,"identity":"883f1cea-7971-444d-991a-ea4a2a2498a6","added_by":"auto","created_at":"2026-05-09 00:04:30","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":96270,"visible":true,"origin":"","legend":"\u003cp\u003eClean and dirty egg (contaminated) detections using the trained (a) YOLO11s \u0026amp; (b) YOLO26s models.\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-9362909/v1/198d04681f044f7e7b7486ee.png"},{"id":108837043,"identity":"38cdda9e-9cb5-4bed-8a88-f92d4cde340e","added_by":"auto","created_at":"2026-05-09 00:04:29","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":262545,"visible":true,"origin":"","legend":"\u003cp\u003eTraining and validation loss of the validation data set for (a)YOLO11s \u0026amp; (b) YOLO26s model.\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-9362909/v1/3e5af16825d3baa506464ba4.png"},{"id":109203398,"identity":"aedefbf8-c1a6-4b20-a1d5-9530dd24dff5","added_by":"auto","created_at":"2026-05-13 14:32:08","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1328009,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9362909/v1/688d06b4-7ce4-4620-8c18-b3d8e3c7cf84.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Performance Evaluation of YOLO11 and YOLO26 for Detection of Low-Contrast Surface Contamination on Eggshell using Fluorescence Imaging","fulltext":[{"header":"Introduction","content":"\u003cp\u003eEnsuring eggshell surface cleanliness is very important for egg and egg products quality compliance as it harbors microorganisms which can further penetrate inside and cause serious food safety issues (Aygun, 2017; De Reu et al., 2006). It can also cause increased mortality of embryos, lower hatchability, and increased early chick mortality (J. Svobodov\u0026aacute;, 2014). Routine inspection under visible illumination can identify gross residues but may miss thin or low-contrast surface films that are difficult to perceive consistently, particularly at industrial speeds which motivating improved, non-destructive sensing methods. In this study, we evaluated fluoresceine imaging as a contrast-enhancement modality for visualizing fecal-smear contamination not readily apparent under standard illumination, and we developed a computer vision workflow for automated detection.\u003c/p\u003e \u003cp\u003eEgg industry is currently undergoing a significant transformation with the adoption of Industry 4.0 technologies, referred to as Egg Industry 4.0 (EI 4.0), which emphasizes automation and digitalization to enhance efficiency and product quality (Ahmed et al., 2023). The main concept of industry 4.0 is digitization of industrial automation system to ensure quality of foods rather than quantity whereas industry 5.0 is mostly focused on sophisticated or highly precise grading considering aesthetic value of food. Traditionally, eggs have been checked visually by product line inspector for dirt and defects, but this approach is slow and prone to error, especially when large numbers of eggs need to be inspected. Some egg industry applied computer vision based visible defect or size and shape-based grading. But some of the dirt on shells remains invisible by naked eye which could compromise the consumer or market standards. Such surface contamination or foreign matters on eggs surface could be visible outside human vision range under UV light and using appropriate optics.\u003c/p\u003e \u003cp\u003eRecently, fluorescence techniques, especially imaging, are getting attention of researchers for potential application in agricultural sciences for quality evaluation (Khaliduzzaman et al., 2023; Konagaya, Omwange, et al., 2020). The techniques involve the detection of naturally occurring fluorescent compounds known as fluorophores which emit light of longer wavelengths upon absorbing energy of shorter wavelengths. FL-imaging provides spatial information on how the fluorophores are distributed and localized in the target sample. This imaging technique provides more specific information than FL-spectroscopic techniques. The most common imaging technique that is used in agricultural sciences is ultraviolet-induced visible fluorescence imaging having excitation wavelength in the UV region and emission in the visible wavelength range (Konagaya, Al Riza, et al., 2020). Thus, the fluorescence imaging of egg surface could visualize this invisible dirt under UV lighting and can be further detected in real time by deploying deep learning object detection like FR-CNN and YOLO. YOLO11 is the latest iteration in the Ultralytics YOLO series for real-time object detection which has proven superior performance over other detectors (Jocher et al., 2023).\u003c/p\u003e \u003cp\u003eYOLO11 introduces significant improvements in architecture and training methods, making it a versatile choice for a wide range of computer vision tasks. These improvements include enhanced features extraction by applying improved backbone and neck architecture, optimized for efficiency and speed up, greater accuracy with lower number of parameters, adaptability across environment for seamlessly deployment and broad range of supported tasks (e.g., instance segmentation, image classification, pose estimation or oriented object detection). Therefore, the aim of this research is to classify clean and dirt eggs non-destructively for industrial grading using UV imaging combined with YOLO11. On the other hand, YOLO26, released in late 2025/early 2026, focuses on extreme deployment efficiency by shifting to an end-to-end, NMS-free design, eliminating the post-processing bottlenecks common in YOLO11 (Jocher \u0026amp; Qiu, 2026). Therefore, fluorescence-assisted imaging combined with deep-learning can support rapid, non-destructive screening for eggshell surface residues and merits further validation using larger datasets that vary by contamination presentation, imaging conditions, and independent production lots.\u003c/p\u003e"},{"header":"Materials and Methods","content":"\u003cp\u003eThis experiment was conducted in accordance with requirements of ethics committee of the Kyoto University, Japan (Approval number: 28\u0026ndash;59).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eImaging system and sample acquisition\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe images of light brown eggs were taken in a dark room using EOS DSLR (kiss x7, Color Inc., Japan) under white and UV light (365 nm LEDs) for color and fluorescence imaging. The EOS DSLR camera was placed 200 mm above the sample. All images were taken with same F-number (f/6.3) and ISO (100). The sample was illuminated using two white bar LEDs (LDL2-80 \u0026times; 16SW2, CCS Inc. Japan) with a shutter speed of 1/5 s for normal color images and two 365 nm UV bar LEDs (LDL-71 \u0026times; 12UV2-365-N, CCS Inc. Japan) with a shutter speed of 2.5 s for fluorescence images of eggs. The distance between the sample and the lighting devices was set as 160 mm (white bar LEDs), and 180 mm (365 nm bar LEDs) for the UV-imaging system. The typical UV images of clean and contaminated eggs are shown in Figure 1.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDataset preparation and annotation\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFluorescence\u0026nbsp;images\u0026nbsp;(\u003cem\u003en\u0026nbsp;\u003c/em\u003e= 74; one image per egg)\u0026nbsp;were annotated\u0026nbsp;in\u0026nbsp;Computer Vision Annotation Tool (CVAT)\u0026nbsp;for dataset preparation. The dataset was split into a 70:30 ratio for training and validation, respectively. The images used for validation were not included in the training set. Two classes (clean and surface contaminated) were defined as Clean_Egg and Dirty_Egg.\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eYOLO detectors and training configuration\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe YOLO11s and YOLO26s models were implemented utilizing the PyTorch framework on a personal computer (12th Gen Intel(R) Core-i7, 2.10 GHz, 64-bit). Some key hyperparameters used during training and inference are summarized in Table 1.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 1.\u0026nbsp;\u003c/strong\u003eHyperparameter configuration for YOLO11s and YOLO26s in eggshell contamination detection.\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 25.4601%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eParameter\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.9509%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eValue\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 58.589%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eDescription\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 25.4601%;\"\u003e\n \u003cp\u003eImage Size\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.9509%;\"\u003e\n \u003cp\u003e640 X 640\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 58.589%;\"\u003e\n \u003cp\u003eInput resolution for training and inference.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 25.4601%;\"\u003e\n \u003cp\u003eBatch Size\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.9509%;\"\u003e\n \u003cp\u003e16\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 58.589%;\"\u003e\n \u003cp\u003eNumber of samples processed before the model is updated.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 25.4601%;\"\u003e\n \u003cp\u003eEpochs\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.9509%;\"\u003e\n \u003cp\u003e100\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 58.589%;\"\u003e\n \u003cp\u003eTotal number of complete passes through the training dataset.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 25.4601%;\"\u003e\n \u003cp\u003eInitial Learning Rate (lr\u003csub\u003e0\u003c/sub\u003e)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.9509%;\"\u003e\n \u003cp\u003e0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 58.589%;\"\u003e\n \u003cp\u003eStep size at the beginning of training.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 25.4601%;\"\u003e\n \u003cp\u003eMomentum\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.9509%;\"\u003e\n \u003cp\u003e0.937\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 58.589%;\"\u003e\n \u003cp\u003eFactor to accelerate gradients in the right direction.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 25.4601%;\"\u003e\n \u003cp\u003eWeight Decay\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.9509%;\"\u003e\n \u003cp\u003e0.0005\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 58.589%;\"\u003e\n \u003cp\u003eRegularization term to prevent overfitting.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 25.4601%;\"\u003e\n \u003cp\u003eOptimizer\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.9509%;\"\u003e\n \u003cp\u003eAuto\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 58.589%;\"\u003e\n \u003cp\u003eAuto (SGD for YOLO11; MuSGD for YOLO26)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 25.4601%;\"\u003e\n \u003cp\u003eBox Loss Gain\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.9509%;\"\u003e\n \u003cp\u003e7.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 58.589%;\"\u003e\n \u003cp\u003eWeighting factor for bounding box regression loss.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 25.4601%;\"\u003e\n \u003cp\u003eIoU Threshold\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.9509%;\"\u003e\n \u003cp\u003e0.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 58.589%;\"\u003e\n \u003cp\u003eIntersection over Union threshold for training anchor assignment.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eYOLO26 removes Distribution Focal Loss (DFL), introduces a redesigned backbone, and uses the MuSGD optimizer to achieve up to 43% faster CPU inference and improved small-object detection. Some basic differences between YOLO11 and YOLO26 are shown in Figure 2.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eYOLO-based inference workflow\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFigure 3 presents the YOLO-based inference pipeline adopted in this study. Images are resized and processed through a convolutional backbone to extract spatial features, which are used to predict object locations and class labels. Final detections are obtained via post-processing, using NMS in YOLO11 and an end-to-end prediction strategy in YOLO26.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ePerformance metrics\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eModel performance was evaluated using precision (P), recall (R), average precision (AP), mean average precision at IoU = 0.50 (mAP50) and mean average precision averaged over IoU thresholds from 0.50 to 0.95 (mAP50\u0026ndash;95) (Khaliduzzaman et al., 2021; Rahman et al., 2025). Precision\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003eis used to measure how well the predictions of the bounding box are correct with the dataset. It is computed by the ratio of actual positive predictions as shown in Eq. 1.\u003c/p\u003e\n\u003cp\u003e\u003cimg src=\"https://myfiles.space/user_files/58895_8739fc6c57c1c19a/58895_custom_files/img1777955565.png\" width=\"598\" height=\"87\"\u003e\u003c/p\u003e\n\u003cp\u003ewhere, TP and FP are known as true positive, false positive values respectively.\u003c/p\u003e\n\u003cp\u003eRecall represents the potential of the model to have the correct prediction of the bounding box measurements within the dataset. It is computed as a ratio of true positive predictions (Eq. 2).\u003c/p\u003e\n\u003cp\u003e\u003cimg src=\"https://myfiles.space/user_files/58895_8739fc6c57c1c19a/58895_custom_files/img1777955606.png\" width=\"573\" height=\"112\"\u003e\u003c/p\u003e\n\u003cp\u003ewhere, TP and FN are known as true positive, false negative values respectively.\u003c/p\u003e\n\u003cp\u003eAverage Precision (AP) summarizes the precision-recall curve for a single class. It is calculated as the area under the precision-recall curve (Eq. 3), integrating precision P(R) as a function of recall, R:\u003c/p\u003e\n\u003cp\u003e\u003cimg src=\"https://myfiles.space/user_files/58895_8739fc6c57c1c19a/58895_custom_files/img1777955638.png\" width=\"510\" height=\"106\"\u003e\u003c/p\u003e\n\u003cp\u003eMean Average Precision (mAP) assesses the overall detection accuracy across all classes. mAP50 is the average precision calculated at a specific Intersection over Union (IoU) threshold of 0.50. It indicates the model\u0026apos;s consistency in detecting objects with moderate overlap. mAP50-95 is the average of mAP values computed at IoU thresholds from 0.50 to 0.95 in steps of 0.05. High mAP50-95 reflects the model\u0026apos;s robustness in precise object localization. The formula for mAP is shown in Eq. 4:\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cimg src=\"https://myfiles.space/user_files/58895_8739fc6c57c1c19a/58895_custom_files/img1777955662.png\" width=\"601\" height=\"112\"\u003e\u003c/p\u003e\n\u003cp\u003e\u003cbr\u003e\u003c/p\u003e"},{"header":"Results and Discussion","content":"\u003cp\u003eAlthough YOLO11s model performed better in terms of accuracy (i.e., precision, recall and thus mAP), YOLO26s model had faster inference time as shown in Table 2. On the validation set, the YOLO11 model achieved higher accuracy of mAP50 = 0.995 and mAP50-95 = 0.995 (precision = 0.908; recall = 0.996) for distinguishing clean shells from fecal-smear contamination whereas YOLO26s model on the same validation set achieved mAP50 = 0.966 and mAP50-95 = 0.966 with precision = 0.834 and recall = 0.946. It was found that YOLO11s was able to detect darts eggs as dart eggs more accurately and it was opposite for the YOLO26s model (detected clean eggs as clean eggs).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 2\u003c/strong\u003e Model evaluation metrics for YOLO11s and YOLO26s validation set\u003c/p\u003e\n\u003cdiv align=\"\"\u003e\n \u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 16.3462%;\"\u003e\n \u003cp\u003eModel\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.3846%;\"\u003e\n \u003cp\u003ePrecision\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.3846%;\"\u003e\n \u003cp\u003eRecall\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 16.3462%;\"\u003e\n \u003cp\u003emAP50\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 16.3462%;\"\u003e\n \u003cp\u003emAP50-95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 20.1923%;\"\u003e\n \u003cp\u003eInference Time (ms)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 16.3462%;\"\u003e\n \u003cp\u003eYOLO11s\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.3846%;\"\u003e\n \u003cp\u003e0.908\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.3846%;\"\u003e\n \u003cp\u003e0.996\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 16.3462%;\"\u003e\n \u003cp\u003e0.995\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 16.3462%;\"\u003e\n \u003cp\u003e0.995\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 20.1923%;\"\u003e\n \u003cp\u003e194.6\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 16.3462%;\"\u003e\n \u003cp\u003eYOLO26s\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.3846%;\"\u003e\n \u003cp\u003e0.834\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.3846%;\"\u003e\n \u003cp\u003e0.946\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 16.3462%;\"\u003e\n \u003cp\u003e0.966\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 16.3462%;\"\u003e\n \u003cp\u003e0.966\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 20.1923%;\"\u003e\n \u003cp\u003e165.3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003eInference time was 194.6ms for YOLO11s whereas 164.3ms for YOLO26s per image on a CPU (Intel i7-12700). The reasons behind the faster inference in CPU are that YOLO26 used Ultralytics edge-optimized model which eliminated NMS with a native end-to-end predictor; removed DFL (Distribution Focal Loss) for simpler, faster inference; introduced MuSGD optimizer (SGD+Muon hybrid) for stable and quick convergence results in faster CPU inference for deployment on low-power devices (Sapkota et al., 2026). These deployment-first design choices align with broader trends in food/agriculture machine vision where inference efficiency and robustness are critical for real-time inspection and sorting pipelines (Shen et al., 2024).\u003c/p\u003e\n\u003cp\u003eMost of the predictions are found to be on the diagonal as indicated by the confusion matrices (Figure 4) implying that the models tend to predict the correct class when a detection is being made. The off-diagonal values indicate the most common practical failure modes, contaminations that are predicted to be clean (false negatives), and clean shells that are predicted to be contaminated (false positives). The off-diagonal counts in YOLO11s are lower than those in YOLO26s, which corresponds to the fact that its precision and recall was higher (Table 2). False negatives tend to be the more expensive error in an inspection process since an egg with contamination can be missed; the increased recall of YOLO11s justifies harsher screening conditions by the conditions of the study. One of the potential reasons for misclassification was that few egg samples contained few shiny dots or white dots (as a contamination) looked very close to clean egg.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe loss curves (Figure 5) give additional reasons as to why YOLO26s is efficient and can be optimized. Throughout training, YOLO26s has smoother and more steady reduction of training and validation losses whereas YOLO11s has stronger spikes in validation losses. Such spikes are common in small or moderately sized datasets, and this suggests that the model is sensitive to a limited number of hard/ambiguous examples (e.g., very thin smears, uneven fluorescence intensity, specular highlights or borderline labeling cases). Notably, YOLO26 is specifically crafted with training-stability techniques, such as ProgLoss and STAL, which are intended to stabilize learning and enhance small/weak-target assignment, as well as MuSGD for better convergence.\u003c/p\u003e\n\u003cp\u003eIn a nutshell, while YOLO11 is a capable, hybrid-task model, YOLO26 is fundamentally redesigned as an edge-first model, focusing on removing bottlenecks for faster, more efficient, and more accurate deployment. One misclassification was due to very few shiny dots or white dots (as contamination) on the eggshell surface which looked very close to clean egg. Hence, egg grader company can consider an omittable threshold to define a clean egg and a surface contaminated egg. Therefore, YOLO26 could be a good choice for real time detection of invisible or low contrast surface contamination of egg at industrial settings.\u0026nbsp;\u003c/p\u003e"},{"header":"Conclusions","content":"\u003cp\u003eEggshell surface contamination can be detected in real-time scenario using fluorescence imaging combined with YOLO11 or YOLO26, as both the models have very good performance on detection of clean and low contrast surface contamination. \u0026nbsp; However, the YOLO26 model might be good for egg industry 4.0 considering faster inference time which is a very important factor for industrial grading application. \u0026nbsp;On the validation set, the YOLO11 model achieved slightly higher accuracy (i.e., mAP50 = 0.995 and mAP50-95 = 0.995) than YOLO26 model for distinguishing clean shells from fecal-smear contamination. On the other hand, YOLO26 performed faster as inference time was 194.6ms for YOLO11 while 164.3ms for YOLO26 per image on a CPU (Intel i7-12700) setting. These results indicated that fluorescence-assisted imaging combined with deep-learning can support rapid, non-destructive screening for eggshell surface residues and merits further validation using larger datasets that vary by contamination presentation, imaging conditions, and independent production lots.\u003c/p\u003e\n\u003cp\u003e\u003cbr\u003e\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eFunding Declaration\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFinancial support was provided by Grant-in-Aid for JSPS Postdoctoral Research Fellow under the JSPS KAKENHI Grant Number 19F19396 and Project Number 82019000018.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics Declaration\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAlthough research manuscripts didn\u0026rsquo;t use any data of any live animal, this experiment was conducted in accordance with requirements of ethics committee of the Kyoto University, Japan (Approval number: 28\u0026ndash;59). Ethics committee (animal research) in Kyoto University follows internationally recognized guidelines for animal experiments.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAlin Khaliduzzaman is on the editorial board member of this journal.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAlin Khaliduzzaman: Conceptualization, Methodology, Investigation, Data curation, Formal analysis, Writing \u0026ndash; original draft. Tanea Gray: Formal analysis, Mamunur Rahman: Writing-review and editing. Naoshi Kondo: Data-curation, Project administration. Isabella Condotta: Writing \u0026ndash; review and editing, Research funding.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eAhmed, M. W., Hossainy, S. J., Khaliduzzaman, A., Emmert, J. L., \u0026amp; Kamruzzaman, M. (2023). Non-destructive optical sensing technologies for advancing the egg industry toward Industry 4.0: A review. In \u003cem\u003eComprehensive Reviews in Food Science and Food Safety\u003c/em\u003e (Vol. 22, Number 6, pp. 4378\u0026ndash;4403). John Wiley and Sons Inc. https://doi.org/10.1111/1541-4337.13227\u003c/li\u003e\n \u003cli\u003eAygun, A. (2017). Chapter 13 - The Eggshell Microbial Activity. In P. Y. Hester (Ed.), \u003cem\u003eEgg Innovations and Strategies for Improvements\u003c/em\u003e (pp. 135\u0026ndash;144). Academic Press. https://doi.org/https://doi.org/10.1016/B978-0-12-800879-9.00013-5\u003c/li\u003e\n \u003cli\u003eDe Reu, K., Grijspeerdt, K., Messens, W., Heyndrickx, M., Uyttendaele, M., Debevere, J., \u0026amp; Herman, L. (2006). Eggshell factors influencing eggshell penetration and whole egg contamination by different bacteria, including Salmonella enteritidis. \u003cem\u003eInternational Journal of Food Microbiology\u003c/em\u003e, \u003cem\u003e112\u003c/em\u003e(3), 253\u0026ndash;260. https://doi.org/https://doi.org/10.1016/j.ijfoodmicro.2006.04.011\u003c/li\u003e\n \u003cli\u003eJ. Svobodov\u0026aacute;, E. T. (2014). FACTORS AFFECTING MICROBIAL CONTAMINATION OF MARKET EGGS: A REVIEW. \u003cem\u003eScientia Agriculturae Bohemica\u003c/em\u003e, \u003cem\u003e45\u003c/em\u003e(4), 226\u0026ndash;237.\u003c/li\u003e\n \u003cli\u003eJocher, G., Chaurasia, A., \u0026amp; Qiu, J. (2023). \u003cem\u003eUltralytics YOLOv8\u003c/em\u003e. https://github.com/ultralytics/ultralytics\u003c/li\u003e\n \u003cli\u003eJocher, G., \u0026amp; Qiu, J. (2026). \u003cem\u003eUltralytics YOLO26\u003c/em\u003e. https://github.com/ultralytics/ultralytics\u003c/li\u003e\n \u003cli\u003eKhaliduzzaman, A., Fujitani, S., Kashimori, A., Suzuki, T., Ogawa, Y., \u0026amp; Kondo, N. (2021). Non-invasive Detection of Chick Embryo Gender Based on Body Motility and a Near-infrared Sensor. In \u003cem\u003eEAEF\u003c/em\u003e (Vol. 14, Number 2).\u003c/li\u003e\n \u003cli\u003eKhaliduzzaman, A., Omwange, K. A., Al Riza, D. F., Konagaya, K., Kamruzzaman, M., Alom, M. S., Gao, T., Saito, Y., \u0026amp; Kondo, N. (2023). Antioxidant assessment of agricultural produce using fluorescence techniques: a review. In \u003cem\u003eCritical Reviews in Food Science and Nutrition\u003c/em\u003e (Vol. 63, Number 19, pp. 3704\u0026ndash;3715). Taylor and Francis Ltd. https://doi.org/10.1080/10408398.2021.1992747\u003c/li\u003e\n \u003cli\u003eKonagaya, K., Al Riza, D. F., Nie, S., Yoneda, M., Hirata, T., Takahashi, N., Kuramoto, M., Ogawa, Y., Suzuki, T., \u0026amp; Kondo, N. (2020). Monitoring mature tomato (red stage) quality during storage using ultraviolet-induced visible fluorescence image. \u003cem\u003ePostharvest Biology and Technology\u003c/em\u003e, \u003cem\u003e160\u003c/em\u003e. https://doi.org/10.1016/j.postharvbio.2019.111031\u003c/li\u003e\n \u003cli\u003eKonagaya, K., Omwange, K. A., Al Riza, D. F., Khaliduzzaman, A., Mart\u0026iacute;nez Oliver, A., Rovira-M\u0026aacute;s, F., Nagasato, H., Ninomiya, K., \u0026amp; Kondo, N. (2020). Association of fruit, pericarp, and epidermis traits with surface autofluorescence in green peppers. \u003cem\u003ePhotochemical and Photobiological Sciences\u003c/em\u003e, \u003cem\u003e19\u003c/em\u003e(12), 1630\u0026ndash;1635. https://doi.org/10.1039/d0pp00236d\u003c/li\u003e\n \u003cli\u003eRahman, M., Souza, V. H. S., Brown-Brandl, T. M., Rohrer, G. A., Shi, Y., \u0026amp; Condotta, I. C. F. S. (2025). 56. Automated monitoring of sow nursing behaviors and postures in farrowing crates through computer vision techniques. \u003cem\u003eAnimal - Science Proceedings\u003c/em\u003e, \u003cem\u003e16\u003c/em\u003e(4), 595\u0026ndash;596. https://doi.org/https://doi.org/10.1016/j.anscip.2025.08.211\u003c/li\u003e\n \u003cli\u003eSapkota, R., Cheppally, R. H., Sharda, A., \u0026amp; Karkee, M. (2026). \u003cem\u003eYOLO26: Key Architectural Enhancements and Performance Benchmarking for Real-Time Object Detection\u003c/em\u003e. http://arxiv.org/abs/2509.25164\u003c/li\u003e\n \u003cli\u003eShen, C., Wang, R., Nawazish, H., Wang, B., Cai, K., \u0026amp; Xu, B. (2024). Machine vision combined with deep learning\u0026ndash;based approaches for food authentication: An integrative review and new insights. In \u003cem\u003eComprehensive Reviews in Food Science and Food Safety\u0026nbsp;\u003c/em\u003e(Vol. 23, Number 6). John Wiley and Sons Inc. https://doi.org/10.1111/1541-4337.70054\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"poultry-science-and-management","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"","sideBox":"Learn more about [Poultry Science and Management](https://poultrysciencemanagement.biomedcentral.com/)","snPcode":"44364","submissionUrl":"https://submission.springernature.com/new-submission/44364/3","title":"Poultry Science and Management","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Eggshell, surface quality, computer vision, object detection, egg industry 4.0","lastPublishedDoi":"10.21203/rs.3.rs-9362909/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9362909/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eEnsuring eggshell surface cleanliness is important for product quality and food safety. Routine inspection under visible illumination can identify gross residues but may miss thin or low-contrast surface films that are difficult to perceive consistently, particularly at industrial speeds. Eggshell surfaces can also carry microorganisms, motivating improved, non-destructive sensing methods. In this study, we evaluated fluorescence imaging as a contrast-enhancement modality for visualizing fecal-smear contamination not readily apparent under standard illumination, and we developed a computer vision workflow for automated detection. Fluorescence images were annotated in CVAT and used to train Ultralytics YOLO11s and YOLO26s object detection models implemented in PyTorch. Images were partitioned into training (70%) and validation (30%) sets. The models\u0026rsquo; performance was assessed on a validation set using precision, recall, and mean average precision (mAP50 and mAP50-95). On the validation set, the YOLO11 model achieved little higher accuracy (mAP50\u0026thinsp;=\u0026thinsp;0.995 and mAP50-95\u0026thinsp;=\u0026thinsp;0.995) but inference time was higher compared to YOLO26 for distinguishing clean shells from fecal-smear contamination under the study\u0026rsquo;s imaging conditions. Inference time was 194.6ms for YOLO11 whereas 164.3ms for YOLO26 per image on a CPU (Intel i7-12700). These results indicate that fluorescence-assisted imaging combined with deep-learning can support rapid, non-destructive screening for eggshell surface residues and merits further validation using larger datasets that vary by contamination presentation, imaging conditions, and independent production lots.\u003c/p\u003e","manuscriptTitle":"Performance Evaluation of YOLO11 and YOLO26 for Detection of Low-Contrast Surface Contamination on Eggshell using Fluorescence Imaging","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-05-09 00:04:25","doi":"10.21203/rs.3.rs-9362909/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"reviewersInvited","content":"","date":"2026-04-24T09:18:21+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-04-10T09:59:35+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-04-10T09:59:21+00:00","index":"","fulltext":""},{"type":"submitted","content":"Poultry Science and Management","date":"2026-04-09T03:54:40+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"poultry-science-and-management","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"","sideBox":"Learn more about [Poultry Science and Management](https://poultrysciencemanagement.biomedcentral.com/)","snPcode":"44364","submissionUrl":"https://submission.springernature.com/new-submission/44364/3","title":"Poultry Science and Management","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"a4918750-b2e0-4cfe-914d-c3e044891ca5","owner":[],"postedDate":"May 9th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-05-09T00:04:26+00:00","versionOfRecord":[],"versionCreatedAt":"2026-05-09 00:04:25","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9362909","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9362909","identity":"rs-9362909","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}