Improved Deeplabv3+ Based Steel Surface Defect Segmentation

Detection of steel surface defects is very important for the steel industry. The occurrence of errors in the production stages of steel also affects the quality of the final product. Therefore, detecting defects on the steel surface is an important step in ensuring the quality and strength of the steel. Traditional steel inspection methods are unreliable in terms of accuracy and consistency and have many disadvantages. Automated inspection systems are promising for the steel industry. In industry, deep learning can help improve defect detection performance and achieve more accurate results. In this study, a new architecture is proposed to improve the segmentation performance of surface defects. In the proposed model, some changes have been made on the DeepLabv3+ model. Thanks to the added Squeeze-and-Excitation (SE) module and the changes made to the Atrous Spatial Pyramid Pooling (ASPP) module, an improvement in the mIoU rate was achieved. The obtained experimental results reveal a significant improvement in the segmentation performance of our proposed model. In comparisons with popular models, the model we recommend has achieved a significant performance improvement.