Research On Classification And Recognition Of Glass Defect Based On Machine Vision

Glass production processes include batching,melting,forming,annealing,processing,and inspection.These processes inevitably carry impurities,leading to defects in the resulting glass.In order to improve the quality of glass,glass defects need to be identified and studied.Traditional manual methods rely heavily on experience and luck,and they are time consuming and inefficient.In order to improve the efficiency and accuracy of defect recognition,this thesis uses the feature extraction technology based on machine vision,and proposes an improved method based on convolutional neural network to achieve accurate identification of glass defects.The main work is as follows:1)Analyze the glass production process and types of glass defects,select four typical defect types and collect various types of images through the optical platform,collect the non-defective images for comparison,and preprocess the various types of images to generate the pictures that meet the experimental needs.2)Feature extraction is used for glass defects.Experiments were performed using the gray-level co-occurrence matrix method.Five feature parameters in the gray-level co-occurrence matrix were selected to determine the step length and direction of the gray-level co-occurrence matrix,and the feature values of all feature parameters of all defects were obtained.At the same time,the moment method was used for experiments and comparative analysis.3)Pattern recognition is used for defect classification.Using convolutional neural network to select appropriate parameters for experiments,the recognition rate of the method is then used K-means method to process the gray level co-occurrence matrix parameters,according to the recognition rate of the two methods proposed improvement program,the improvement program is to use K-means method corrects the output values of the convolutional network to achieve optimization of the convolutional neural network classification.The experimental results show that the improved network model can correct the eigenvalues of the original erroneously sampled samples to a certain extent,and at the same time,the eigenvalues of the original categorized correct samples can be preserved,so that the final recognition rate reaches 95.71%.