DA-UNet: An improved UNet-based algorithm for segmenting defects in printed cards

Many deep learning-based methods have been widely applied in the field of printed defect detection. However, most methods still face challenges such as misclassification, weak ability to detect multi-scale defects, and complexity in model architectures. Therefore, we propose an improved UNet-based defect segmentation network named DA-UNet. In the backbone of the network, DA-UNet utilizes Dilated Asymmetric Convolution Modules (DACM) to replace regular convolutions, reducing parameter count while enhancing the network’s ability to represent multi-scale features, thereby improving its capability to detect multi-scale defects. During the fusion process of encoder and decoder features, an Adaptive Feature Fusion Module (AFFM) is employed to learn the weights of corresponding positions of features, enhancing the network’s ability to focus on key features. The algorithm achieves an average Intersection over Union (IoU) of 92.26% and an F1 score of 92.90% on the printed card defect dataset, with lower algorithmic complexity and faster inference speed. Visualization results indicate that the algorithm maintains high segmentation accuracy under different lighting conditions, exhibiting strong robustness, and suitability for defect segmentation tasks in industrial printing environments.