Defect Segmentation of Magnetic Tiles with the Novel Ardise-U-Net

Detecting and diagnosing surface defects is critical throughout the production process. The detection of these defects with high precision plays a major role in preventing both financial and temporal losses. Recently, image processing, machine learning, and deep learning-based approaches have made great progress in surface defect detection. Therefore, in this paper, a new approach called Attention Residual Dilation Squeeze and Excitation U-net (ARdiSE-U-net) is proposed for surface defect detection. The suggested method includes an encoder and decoder, which are the key components of the Unet structure, a depth-wise compression and stimulation block summed to the Unet's skip connections, and a hybridization of attention residual blocks. A model created by applying multiple improvements on the classical U-Net architecture is described. The model integrates attention mechanisms, residual blocks, dilation applications, and Compression-Evoked (SE) blocks. This combination aimed to improve the segmentation performance and provide deeper feature extraction. In this way, critical information is extracted. Lastly, pixel-stage flaw detection is designed with the help of the sigmoid function in the output layer. In the developed model, MT dataset is used for magnetic surface flaw detection. In the experimental studies, Accuracy, Mean IoU, AUC, MAE results of the developed ARdiSE-U-net architecture are 0.9893, 0.8128, 0.8936, 0.0099, respectively. It has higher performance metrics than most of the baseline studies.