In summary, the contributions of our study are listed below.
• The first to adopt the YOLOF model on a metal defect NEU-DET dataset.
• Using several image preprocessing methods, including Edge Detection, Denoising, Sharpening, and SR, were experimented on the metal defect NEU-DET dataset.
• In the proposed optimized batch size, backbone, soft efficient non-maximum sup-pression (Soft-NMS), and IoU threshold of YOLOF; we call it the YOLOF+ model, and it shows some improvement.
• Experiments Super Resolution - BSRGAN for patches and Sharpening for rolled in scale have better results than the YOLOF+ model for patches and rolled in scale.