TD‐YOLO: A Lightweight Detection Algorithm for Tiny Defects in High‐Resolution PCBs

• DOI: 10.1002/adts.202300971
• Publication Date: 2023-12-23T04:26:03Z
• Authors: Qin Ling; Nor Ashidi Mat Isa
• Keywords: Pyramid (geometry); Feature (linguistics); Neuromorphic engineering

Abstract Printing circuit board (PCB) defect inspection precisely and efficiently is an essential challenging issue. Therefore, based on several improvements upon YOLOv5‐nano, a novel lightweight detector named TD‐YOLO proposed to inspect tiny defects in PCBs. First, the ShuffleNet block implemented into backbone effectively reduce model weight. Second, anchors are designed using modified k‐means clustering accelerate convergence yield superior detection precision. Then, data augmentation strategy recomposed by rejecting mosaic suppress emergence of extremely targets. Finally, mighty feature pyramid network namely MPANet, newly boost fusion capability model. The experiment results denote achieves highest 99.5% mean average precision our dataset, outperforming other state arts. Specially, metrics for smallest two defects, such as spur mouse bite, increased 2.1% 1.2%, respectively, compared with YOLOv5‐nano. Besides, has only 1.33 million parameters, decreased 25% than baseline. Using mediocre processor, speed boosted 20%, reaching 37 frames per second input size 22402240 pixels.