Object Detection Based on Enhancement of Subtle Features in Low-Illumination Images

Abstract This study discusses the issue of target detection in low-light environments, which is of great significance for improving traffic safety, enhancing the perception capabilities of autonomous driving, and intelligent surveillance systems. Addressing the challenges of image detection under low-light conditions, such as low brightness, weak contrast, and noise interference, this paper selects the YOLOv5 model for improvement research and primarily proposes a new type of attention mechanism, ICA, which integrates IRMB and CA attention mechanisms to more accurately identify inconspicuous features and reduce the interference of irrelevant information. Additionally, to address contrast issues caused by uneven lighting, this paper replaces the original YOLOv5 up-sampling operator with the CARAFE operator, which dynamically generates adaptive up-sampling kernels through content-aware methods and aggregates contextual information within a larger receptive field, thereby enhancing the model's object recognition capabilities. Lastly, to cope with the variability of objects in low-light environments, this paper introduces the ASFF module, which further enhances the model's detection performance for multi-scale targets by adaptively learning the fusion weights of each scale feature map. Through multiple training sessions, more suitable parameter settings for training are found, further enhancing the model's performance. The proposed model is named YOLOV5-ICAs.