Fast recognition method of surface defects on micro-steel balls based on improved AlexNet-SVM convolutional neural network

The quality control of surface defects on micro-steel balls is particularly difficult because of high reflectivity and the need for full coverage on the spherical surface. In response to the problems of low efficiency and inadequate accuracy of manual inspection methods, an improved method is proposed for fast identification of surface defects on steel balls with a combination of an improved AlexNet convolutional neural network and an SVM model. The last three convolutional layers are removed and the features extracted by the fully connected layer FC7 are reserved in our model. The original Softmax classifier is replaced by SVM to prevent overfitting and improve the model's generalization ability. In addition, an improved network search algorithm based on K-CV is used to determine optimal parameters for the classifier. Experimental evaluation of the proposed model's recognition results is performed with a confusion matrix. The results show that this method achieves an average accuracy rate of 99.43% with an operating time of 17.2 ms. Compared to the original model and other network models, it has higher accuracy and inference speed, meeting the requirements of industrial field inspection.