DesNet: PCB defect detection network based on deformable convolution

Printed circuit boards (PCB) defect detection is very important in today's developed manufacturing industry, the surface defects of a PCB not only affect the aesthetics and comfort of the product, but generally also have adverse effects on its usability. Traditional inspection methods rely on the intuitive visual judgement of the operator, with strong subjective judgement factors, low accuracy rates and high staff costs. Deep learning can be a good solution to this problem to reduce the investment in human and material resources. A method/network structure for PCB defect detection based on deformable convolution has been proposed. The aim is to address the problems associated with existing PCB defect detection algorithms, which include low detection accuracy, poor real-time performance, and insufficient robustness. The network depth is increased to extract more and higher-level semantic information, which is guaranteed to be gradient-free under the resnet feature improved. The improved network model has demonstrated enhanced performance in the field of PCB defect detection. Specifically, it has outperformed the original RetinaNet network model by achieving a 3% increase in mean accuracy(mAP) on the PCB defect detection dataset. Additionally, the detection speed of the modified model satisfies real-time requirements, further demonstrating its practical utility.