Research on a Deep Learning-Based Hybrid Attention and Regularised Supervised Detection Model for Surface Defects in Precision Components

To solve the challenges in detecting hundreds of minute surface defects (scratches, pits, material shortages) and their combinations on precision components (connecting rods, gears, bearings, etc.) ranging from 6 to 30 mm in size—notably identification difficulty and high misdetection rates—this research proposed a deep learning convolutional neural network detection model integrating a with dropout regularization supervision. This model aims to achieve rapid identification and precise judgment of micro-defects, combined defects, and other issues in precision components. This model extracted global fragmented features from precision component images through convolutional layers. Pooling layers incorporated a hybrid attention mechanism integrating maximum and average pooling. Fully-connected layers concurrently assembled hybrid attention dual-pooling filtered features, creating hybrid attention weights via hyperbolic tangent activation functions to obtain precise fusion of location feature and defect type extraction. The dropout regularization mechanism randomly deactivates relevant neurons during the feature fusion stage of the fully connected layer. Experimental results demonstrated that this model reduced false detection rate for similar defects (e.g., dents of varying shapes and scratches with different contours) to obtaining an average defect detection accuracy of 99.8%. with an average detection speed of 300 pcs/min. This represented improvements of 9.2% and 7.6% in detection accuracy and 15% and 10% in detection speed compared to traditional CNN and single-attention models, respectively. It also presented excellent robustness in random defect detection under field conditions, providing a novel solution for efficient and precise surface defect detection on precision components.