An Improved YOLOv8 Network for Photosensitive Element Defect Detection

The photosensitive element plays a critical role in ensuring high image quality within camera modules. Detecting Dead Pixel and stain defects during production is crucial for maintaining this quality. However, the current target detection algorithm faces several challenges in accurately detecting these defects. Firstly, the small size and limited image information of Dead Pixel defects make it difficult to extract accurate features for classification and positioning. Secondly, label production for small Dead Pixel defects and unclear stain defect features often leads to missed or mislabeled defects, resulting in low-quality samples and unsatisfactory regression results. Lastly, the varying sizes of stain defects pose difficulties in detection, leading to suboptimal performance. To address these challenges, we propose an accurate defect detection model for photosensitive elements based on an improved version of YOLOv8. Specifically, we integrate the CBAM attention mechanism into the convolutional layer to focus on small defects and suppress background noise, effectively extracting feature information. Moreover, we employ the WIoU loss function to improve regression on regular samples while reducing attention to low-quality labels. Additionally, we replace the SPPF module with the SimCSPSPPF module, expanding the convolution branch and reducing computation, thereby improving the detection of larger targets like stain defects and increasing inference speed. Experimental results demonstrate the superiority of our method over multiple representative methods. Compared to the second-best method (YOLOv8), our method achieves significant improvements in the average precision of Dead Pixel defects (86.9% to 88.0%), the average precision of stain defects (93.3% to 94.7%), and the average accuracy (90.1% to 91.3%).