A Dense Sample Surface Defect Detection Algorithm Based on Deep Learning and Thread Pool

Dense samples are prevalent in various fields, including biopharmaceuticals, semiconductors and component manufacturing which are characterized by tiny size, large quantities and tightly packed. In practical quality control, multitude of samples and individual variances, influenced by material properties, processes et al, pose substantial challenges for precise and rapid surface detection. This paper proposes a dense sample surface defect detection algorithm based on precise localization of pre-selected features and scalable thread pool. The algorithm constructs a lightweight network model by adjusting the convolution depth and the count of CSPNet (Cross Stage Partial Network) structures in the YOLOv4-tiny to achieve efficient and accurate positioning of individual samples. A thread pool with a scalable task queue is designed to improve detection efficiency. Experimental results show that the mAP (Mean Average Precision) of sample detection is as high as 99.35% under the IOU (Intersection over Union) threshold of 0.91. Furthermore, with this algorithm, the detection efficiency will be improved by about 20% in dense sample cases.