Union channel pruning-based U2Net for online surface defect segmentation of aluminum strips in production processes

The automated defect inspection (ADI) of aluminum strip surfaces encounters several issues in the practical production process, such as the challenge of achieving precise defect boundary identification and the high demands for model volume and real-time inference. To solve these problems, the inference process of Input-Conv-BN-ReLU-Output in every layer of the high-precision segmentation model U2Net is fully analyzed and a novel union channel pruning (UCP) algorithm based on the U2Net is designed to significantly simplify the model structure. The absolute values of the convolution weights in the channel are first added up as the first indicator. Then the expectation of the truncated Gaussian distribution that processed by the batch normalization (BN) layer and ReLU activation layer is calculated as the second indicator because it makes reasonable use of the scale factor, shift factor, and interval information. The two indicators are multiplied as the final evaluation indicator, which realizes the comprehensive consideration of the Input-Conv-BN-ReLU-Output in U2Net. Additionally, we collect the surface images of aluminum strips from the online inspection platform and create a new dataset with seven common defects. Experimental findings obtained on the dataset demonstrate that the UCP performs better than other network slimming approaches, especially at high pruning ratios. The U2Net with 62.5% of channels pruned by the UCP method surpasses other cutting-edge and lightweight segmentation models in segmentation accuracy and speed, which may serve as a valuable theoretical guidance for the automated online defect segmentation of aluminum strips on embedded devices.