Steel surface defect detection based on PSO-Gabor and an improved faster RCNN

Thin strip steel, as a primary product in the metallurgical industry, is extensively utilised across various fields, so it is crucial to carry out non-destructive testing of its defects during production. Existing datasets, such as NEU-DET, are incomplete and the contrast of images differs from the actual production. Therefore, combined with NEU-DET, a new dataset is constructed by adding self-built hole and edge crack defect images. Additionally, a defect detection algorithm based on image enhancement and deep learning was proposed. The PSO-Gabor algorithm enhanced the contrast of defect images and highlighted the defect features. An optimised Faster RCNN network was proposed, which used the improved AlexNet as the feature extraction network. The SE attention and the depthwise convolution were introduced to capture local features.an IoU K-means improved the accuracy of anchor box generation. The pooling layer was replaced to address the feature maps size deviations. Finally, a defect detection platform was designed to conduct an experimental validation. The results showed that the improved AlexNet network achieved a classification accuracy of 99.78% for surface defects on steel plates, which was a 17.9% improvement over that of traditional networks. The optimised Faster RCNN achieved a mAP improvement of 0.14 while maintaining an FPS of 35.