SMD LED chips defect detection using a YOLOV3-dense model

This research develops a defect detector on the basis of the you only look once version 3, YOLOv3, and applies a process to inspect the chips of surface-mounted device light-emitting diodes (SMD LED). In addition to an inspection process that identifies defect-free products, the model also identifies missing components, incorrect placement, inverse polarity, missing wires and defective surface. To improve the performance of the defect detector, the research substitutes densely connected convolutional networks (DenseNet) for the Darknet-53 backbone, producing a model that is called YOLOv3-dense. The research introduces the Taguchi method to evaluate the sensitivity of YOLOv3-dense's hyper-parameters in a fewer number of experiments. The result shows that the testing mean average precision (mAP) measured by YOLOv3-dense is 33.69% higher than the class activation mapping (CAM) localization of convolutional neural network (CNN). The testing mAP is also 14.98% higher than traditional YOLOv3.