Semantic Segmentation based Defect Detection Algorithm for Ceramic Substrates

Semantic segmentation, an important technique, has been widely applied in defect detection and recognition in automated systems. However, it still faces challenges such as high parameter count and inaccurate defect boundary localization. To address these issues, this paper proposes a lightweight semantic segmentation algorithm for defect detection based on DeepLabV3+. Firstly, we utilize a lightweight network, MobileNetV3, to extract defect features from ceramic substrates. Secondly, we employ a parallelized approach with multiple layers of dilated separable convolutions to further extract deep features, which not only reduces the parameter count but also expands the effective receptive field. Lastly, we fuse the deep features, which incorporate essential channel features, with shallow features, and subsequently upsample the fused features using bilinear interpolation to generate the final segmentation predictions. Experimental results, when applied to defect detection in ceramic substrates, demonstrate that the improved model achieves higher segmentation accuracy, utilizes fewer parameters, and improves defect boundary localization capability.