Small Object Detection of Non-standard THT Solder Joints Based on Improved YOLOv3.

Object detection on non-standard THT (Through Hole Technology) solder joints is a difficult task, due to the reason that solder joints are small targets, which are hard to detect on PCB (Printed Circuit Board). Moreover, there is a lack of datasets and benchmarks for non-standard THT solder joints detection. To solve the two issues mentioned above, we propose a new object detector (YOLOv3_5L) and establish a new dataset of non-standard THT solder joints. Based on YOLOv3, the YOLOv3_5L adopts five scales of feature maps instead of three scales of feature maps to predict bounding boxes of objects, which can obtain more texture and contour information to detect small objects. Extensive experiments on our dataset show that YOLOv3_5L have better performance than YOLOv3 while ensuring that the speed is almost unchanged. On our test set of $1024\times 1024$ pixels, the mAP@0.5 result of YOLOv3_5L is 97.2%, which improves about 1.2% compared with baseline model (YOLOv3).