Small target detection of bearing surface defects based on improved YOLOv5s

Abstract

In the context of industrial automation, the accurate detection of small defects on bearing surfaces (dents, bruise, scratch) is crucial for the safe operation of equipment. However, traditional detection methods have problems such as insufficient feature extraction for small targets and sensitivity to background interference. Based on this, an improved YOLOv5s small target detection method ECN-YOLOv5s for bearing surface defects is proposed. First, an efficient channel attention mechanism (ECA) is inserted in the layer before the SPPF in the backbone network, which realizes efficient computation of channel attention and reduces redundant computation while decreasing the model complication. Second, using the Content-Aware ReAssembly of Features (CARAFE) module instead of the original nearest-neighbor up-sampling module achieves light weighting while allowing for better aggregation of contextual information within a larger sensory field, which effectively improves the diversity and effectiveness of the model. Finally, an original loss function is replaced with Normalized Wasserstein Distance (NWD) to decrease the susceptibility to small object positional deviations. Experimental results in our homemade dataset indicate that the mean average precision (mAP) of ECN-YOLOv5s for bearing defect identification is 92.7%. This is an improvement of 2.4% versus the former YOLOv5s model. Compared with some mainstream detection models, the modified model in this study also provides better effects. Thus, the model is able to meet the requirements for the detection of small defects in bearings in industry.