Convolutional Attention Network with edge supervision for void and spot segmentation in chip resistors based on X-ray images

Abstract

The presence of voids inside solder joints in chip resistors can affect the stability and reliability of the product. Therefore, accurate void and spot segmentation is crucial for performance evaluation. In this study, we propose Convolutional Attention Network (CA-net) with edge supervision for void and spot segmentation based on X-ray images of chip resistors. First, CA-net has an encoder-decoder architecture and use skip connections to aggregate low-level features and high-level features. Second, the multi-scale convolutional attention modules (MSCA) are employed at each encoder and decoder to capture multi-scale feature maps of voids and spots. Finally, adding edge-supervised branch to the encoder allows the network to focus more on edge information, so that the response area of the network will be more focused on pads and empty areas. Experimental results show that CA-net outperforms other methods in void and spot segmentation, achieving 96.56 % MIOU and 97.44 % mean DIC. These results indicate that CA-net is a promising approach for accurate void and spot segmentation in chip resistors based on X-ray images.