SSIM-NET: Real-Time PCB Defect Detection Based on SSIM and MobileNet-V3

Defect detection is an indispensable part of the print circuit board (PCB) manufacturing process, which could identify problems in the previous manufacturing process to avoid the subsequent meaningless assembly work. However, previous works, e.g., traditional detection approaches fully based on machine learning, which are really too sensitive by some environmental factors. In this work, we improve previous work and propose a new PCB defect detector called SSIM-NET. Our approach is a two-stage detection algorithm including two innovations: 1) In the first step, we use the structural similarity index (SSIM) instead of morphology operator to find out the suspicious regions, it is not easily susceptible to some factors, e.g., illumination variation, slight camera noise, etc. 2) In the second step, in order to reduce the computational cost, we adopt the latest lightweight backbone MobileNet-V3 to classify suspicious regions. The difference is that we add a binary loss that focal loss to accelerate the network coverage. Finally, the accuracy and speed of our approach both achieved the state-of-the-art. Specially, compared with Faster-RCNN, its speed is faster at least 12× but without any loss of accuracy.