Bearing surface defect detection algorithm based on deep learning

Taking into account the industrial environment, where target detection algorithms aimed at tiny defects on the bearing surface suffer from low precision and slow speed, an improved RT-DETR algorithm, based on deep learning techniques is proposed. Firstly, the FasterNet is utilized to rebuild the backbone network, reducing computational stress while embedding a DWRF module integrated with EffectiveSE attention mechanism, allowing for in-depth feature extraction and bolstering finer-grained feature depiction, consequently reducing missed detections. Secondly, by incorporating Cascaded Group Attention (CGA) and designing an S-Fusion module embedded with SimAM attention, further limit computational redundancy and amplify the ability to capture key features. Finally, the NWD (normalized Wasserstein distance) combined with Inner-MPDIoU is applied for joint optimization loss, hastening the model's convergence and boosting the detection accuracy for small targets. The experimental results illustrate that in comparison to the original RT-DETR algorithm, the improved algorithm increases the mean average precision by 2.4%, reduces computation by 16.1%, and augments detection speed by 15.6%. Furthermore compared with Faster R-CNN, YOLOv5, YOLOv8 and other methods, our algorithm exhibits superior detection speed and accuracy, thereby meeting the requirements for detection of bearing surface tiny defects in industrial environment.