Research On Glass Defect Recognition Method Based On Improved Convolutional Neural Network

In the process of glass production,defects such as scratches,inclusions,bubbles,tumors and stains may occur.There is no fixed pattern in the shape and size of various glass defects.The method of manually selecting features for defect identification and classification can only be completed for a specific glass or defect,the classification is time consuming and accuracy is low.Therefore,finding a versatility,high accuracy and high efficiency of glass defect identification classification method is of great significance for improving the quality of glass products.In this paper,dual cold cathode fluorescent lamp is used as the source of linear CCD camera to detect the defects of the glass products on the glass production line.For the strip noise caused by the stroboscopic characteristics of the light source in the image,this paper designs a band-stop filter to eliminate,enhances the image by grayscale piecewise linear transformation,thus improves the contrast between the glass defect and the background neighborhood.Aiming at solving the problem of poor versatility of artificially extracted image features in traditional classification methods,this paper uses the method of glass defect recognition based on convolutional neural network.Based on the analysis of deep learning theory,geometric transformation is used to expand the dataset of glass defect image,and the convolutional neural network structure suitable for glass defect recognition was determined by trial and error.In the process of convolutional neural network training,the convolution kernel is randomly initialized and trained,resulting in low network efficiency.By combining supervised and unsupervised learning,an integrated learning method is proposed in this paper.The method uses the sparse self-encoder with improved sparse coefficient to learn the hidden features of image blocks,and takes the learned features as the initial convolution kernel of the convolutional neural network,which effectively reduces the training time and by replacing the Softmax function with an improved KSVD to identify defects,the recognition accuracy is improved.In order to further improve the recognition accuracy of the integrated convolutional neural network for inclusion and tumor type defects,a multi-channel integrated convolutional neural network is constructed.The network combines multiple single-channel integrated convolutional neural networks to construct a multi-channel integrated convolutional neural network,and the method of combining multiple single-channel networks is given.Through the defect enhancement of different methods of glass defect images,different network channels can learn the defect images with different contrasts,thus obtaining a deeper semantic description of more complete and accurate glass defect images.Through the performance comparison,the optimal channel number of multi-channel integrated convolutional neural network is determined,and the best multi-channel integrated convolutional neural network is used to identify the glass defects,which effectively improves the misidentification between the inclusions and the tumor type defects and improves the recognition accuracy of glass defects.