Perception of Flaws in Steel using ResNet50 Algorithm

Automatic identification of steel surface faults is critical in the steel industry for ensuring product quality. The old technique, however, cannot be employed efficiently in a manufacturing line due to its poor accuracy and slow working speed. The existing, well-liked deep learning-based algorithm also suffers from a lack of precision, and there is still much opportunity for development. This paper presents an approach for reducing average running time while enhancing accuracy by combining upgraded faster region convolutional neural networks (faster R-CNN) with improved ResNet50. The updated ResNet50 model starts with a picture and then adds the deformable revolution network and a better cutoff to determine if the sample has defects or not. The algorithm outputs a sample free of flaws if the likelihood of a defect is less than 0.3. Otherwise, the data are added to the modified faster R-CNN, which includes matrix NMS, enhanced feature pyramid networks, and spatial pyramid pooling. The location and classification of the sample defect, if present, are the output's final results.