FFDR-Net: Feature Fusion Deeper RetinaNet for PCB Defect Detection

Fast and accurate detection for small defects in printed circuit board processing helps improve industrial efficiency and expand industry profits. The advancement of deep learning technology has dramatically increased the speed and precision of object detection in natural scenes, and it also promotes the development of the defect detection field. Due to the irregular shape and small size of circuit board defects, existing defect detection methods have problems such as low detection accuracy. In this research, we suggest a defect detection network based on RetinaNet network named FFDR-Net. The network is based on the framework of RetinaNet model, and ResNet-152 with deeper network layers is employed as the backbone structure to enhance the network feature processing capability; Meanwhile, the PaFPN structure is used in the neck part to improve the detection accuracy of small-sized objects by including a bottom-up feature fusion module, which reduces the information flow between the lower and the top layer features; In addition, a new data enhancement strategy is added during training to improve the object localization accuracy by randomly adjusting the image chromaticity, saturation and brightness. Finally, ablation experiments were conducted on the PCB dataset, and the revised network model increased mAP by 6.6% compared with the original RetinaNet model, and the average accuracy AP of different object class detection was mostly improved, which accomplished the recognition task better in the actual detection process.