Defect detection and identification model of mechanical manufacturing products based on machine vision and deep learning

This paper introduces a defect detection model of mechanical manufacturing products based on machine vision and deep learning (DL), which includes five modules: image acquisition, preprocessing, feature extraction, classification recognition and result output. By using high-resolution camera and optimized convolutional neural network (CNN) architecture (adding 2-3 convolution-pooling layers on the basis of LeNet-5), clear product surface images are collected and preprocessed, and the key features in the images are automatically learned and extracted. In the experiment, the learning strategy of 3×3 convolution kernel, 2×2 maximum pooling, initial learning rate of 0.001 and batch size training model of 32 were used for 100 rounds. The model shows more than 90% accuracy, recall and F1 score in defect detection of bearings, gears and bolts, and is especially good at identifying cracks, depressions and notches. Even under different lighting conditions and surface pollution, the model still performs well, which proves that it has good robustness and generalization ability and is suitable for defect detection in production line.