YOLOv8-SST Lightweight Small Target Fault Detection System Integrating SENet Attention Mechanism and Tr-OCR

Abstract To solve the issues of slow processing speed and susceptibility to false and missed detections in fault detection algorithms with cluttered backgrounds, this paper proposes an improved YOLOv8 lightweight fault small target detection method based on the fusion of a squeeze-and-excitation network and transformer-based optical character recognition. The backbone network structure of the YOLOv8 algorithm was optimized by adding the channel attention module SENet to enhance the performance, adaptability, and generalizability of the model. The differentiable feature learning loss function was optimized to guide learning, reduce positioning errors, and improve accuracy. Group shuffle convolution was used in the neck layer to replace the traditional standard convolution to construct a lightweight detection network, which increases the diversity of the network receptive fields, better captures contextual information regarding the target, and reduces complexity. This study innovatively introduced optical character recognition (OCR) into the YOLOv8 backbone to improve digit recognition accuracy. A CCD-type image sensor was used as an image acquisition device, to further improve the accuracy of the fault detection. Finally, the experimental results were analyzed, and the improved YOLOv8n algorithm achieved a detection precision of 96.84%, a recall of 98.73%, and an F1-Score of 97.81%.