Real-time steel surface defect detection using RT-LSM-DETR

During steel production, surface imperfections often emerge, compromising product quality. This study introduces RTLSM-DETR, a novel real-time detection network designed to overcome the limitations of conventional methods, which suffer from poor accuracy and slow processing speeds. Building upon the RT-DETR framework, this network incorporates various modules to substantially improve its ability to identify steel surface flaws. The original model's Cross-scale Feature Fusion Module (CCFM) utilizes fusion blocks arranged in convolutional layers to efficiently combine adjacent features, thus enhancing feature representation. To further refine the CCFM, Large Separable Kernel Attention (LSKA) is implemented in place of RepC3 convolution. LSKA breaks down large convolutional kernels into a series of smaller ones through separation, thereby broadening the receptive field and boosting the network's capacity to capture global information without significantly increasing computational demands. Additionally, the Multi-Scale Convolutional Attention (MSCA) module is incorporated, employing multi-scale convolutions to detect various feature scales and utilizing attention mechanisms to highlight crucial features. This integration enhances the network's responsiveness to steel defects and improves detection precision. Experimental outcomes demonstrate that the enhanced RT-LSM-DETR network model exhibits superior detection performance, achieving a mAP of 94.0% on the NEU-DET dataset with MixUp data augmentation, representing a 2.9% improvement over the original RT-DETR network. Furthermore, the detection speed reaches 102.04 FPS, meeting the requirements for real-time steel defect detection in industrial applications