YOLOv7-BA: A Metal Surface Defect Detection Model Based On Dynamic Sparse Sampling And Adaptive Spatial Feature Fusion

The detection of metal surface defects is crucial as metal materials play a vital role in national construction and are widely used in various industrial manufacturing fields. However, differences between the types of metal defects can be subtle, which makes it challenging for traditional coarse-grained detection methods to achieve satisfying detection accuracy. Therefore, in our research, a new model YOLOv7-BA is proposed based on dynamic sparse sampling and adaptive spatial feature fusion. The former can extract the fine-grained features, while the latter can leverage the coarse-grained information in the meantime. Thus, it enables the model to distinguish subtle differences between different defect types. Extensive experiments are conducted to evaluate the effectiveness of the proposed YOLOv7-BA against other detection models on the NEU-DET metal surface defect dataset from Northeastern University. Results indicate that YOLOv7-BA achieves an average detection accuracy of 74.8%, which is 3.6% higher than that of the YOLOv7 model. The accuracy concerning patches and pitted surfaces is 91% and 96.6%, respectively. YOLOv7-BA demonstrates high precision in detecting small target defects and exhibits strong real-time performance.