Efficient Faster R-CNN: Used in PCB Solder Joint Defects and Components Detection

Abstract

In this paper, an improved efficient Faster R-CNN algorithm used in PCB solder joint defects and components detection is proposed. The accurate detection of solder joint defects and components on printed circuit boards (PCB) is very important to the quality of production of electronic products. The traditional automatic optical inspection needs a customized registration and program design for each circuit board. The necessary steps are complex and time-consuming, and the judgment standard is relatively simple and rigid, resulting in a poor generalization ability of detection and decreased detection accuracy, especially in the conditions of poor lighting, color change of circuit board, and multiple-variety and mini-batch production scenarios. In this paper, an improved Faster R-CNN PCB assembly solder joint defects and components detection algorithm, named Efficient Faster R-CNN, is proposed based on EfficientNet-B7. In this algorithm, the first eight stages of EfficientNet-B7 network structure are used to replace VGG-16 for feature extraction, and the generalized intersection over union ratio is used as the loss of bounding-box regression to avoid the non-optimal solution caused when the prediction boxes and the real boxes are disjoint. The proposed algorithm also uses the swish activation function with better performance to replace ReLU.The experimental results show that the detection accuracy of the proposed algorithm is significantly improved, with the mean average precision close to 0.99, which is of great significance for industrial application.