Optimize D-LinkNet for Printed Circuit Board Defects Inspection

Abstract

According to Taiwan Printed Circuit Association (TPCA) statistics from 2017 to 2021, the output value of PCBs in Taiwan has increased year by year or even broken through new highs, and the cost of labor and time to visually inspect scratches on PABA by personnel has increased relatively. Therefore, this study focuses on PCBA for scratch detection and trains multiple models based on semantic segmentation UNET network architecture. The proposed D-LinkNet is optimized to reduce the problem of missed detection and misclassification caused by complex backgrounds and long span of defects. By comparing various attention modules in different positions and types to improve the accuracy, and using the dilated convolution instead of pooling layer, the encoder-decoder structure is optimized to reduce the loss of information in the downsampling process, simultaneously improve attention module effect. In addition, this experiment uses a small amount of data to increase the amount of data by cutting and augmenting the data, and compares the effect of image cutting size on the accuracy rate to find the best data size for training, and uses IoU as the model scoring method to apply the model with the best segmentation effect to more scratch detection tasks and reduce the labor cost at the factory.