Defect detection is an essential and, moreover, challenging procedure for enhancing the quality of industrial products. To implement an intelligent defect detection process using real‐time image processors based on deep learning, some important points should be considered. Therefore, in this paper, three levels will be mentioned; first, in the smart production lines of industrial elements, the processor must classify whether the elements have defects or not. By designing the Inception40 + Vgg16 network and applying it to the entire dataset, defect detection is achieved with an accuracy of 98. 6% ± 0. 8% (fivefold cross‐validation). Second, the processor must be able to classify the type of the defective elements, which is an important part for the defect inspection, and consequently improve the efficiency of production quality of the elements. By applying Dense27 + Vgg16 classification network to defective images, the defect type of images is classified with an accuracy of 97. 4% ± 0. 9%. Finally, the defect will be detected in the last step. At this step, each image goes through the contrast limit adaptive histogram equalization (CLAHE) filter for adjusting, with the implementation of the UD52 + DeepCNN segmentation network. Our approach addresses the entire defect detection pipeline with systematically designed network architecture grounded in transfer learning theory and receptive field analysis, achieving segmentation performance of 81. 78% ± 1. 3% and 80. 01% ± 1. 5% MIoU on DAGM 2007 and road defect datasets, respectively. Statistical significance testing confirms the superiority of our approach over baseline methods (p < 0. 05 for all comparisons). Comprehensive ablation studies quantify the contribution of each architectural component, and domain shift analysis validates cross‐domain transferability. Our codes are available at https: //github. com/Alirezanltv/Patchbaseddefectₛegmentation.
Kanani et al. (Thu,) studied this question.