Research On Workpiece Recognition Technology Based On Convolutional Neural Networks

For the traditional workpiece recognition process is designed by manual according to the characteristics of the workpiece features to be extracted,and the use of simple algorithms to complete the whole process has a high time-consuming,costly,poor generality and recognition accuracy is not high,this paper combines machine vision technology and convolutional neural network algorithm for workpiece recognition and classification,the system can complete automated image feature acquisition,and classification The system can complete automatic image feature acquisition,and classification and recognition process into one,and improve the workpiece recognition accuracy and recognition speed,the main work content is as follows.（1）For the convolutional neural network model requires a large amount of data for training,the original data of eight types of workpieces collected were subjected to data enhancement operations and divided into three data sets according to the actual industrial production scenarios;in order to improve the image quality of the data sets,a series of pre-processing operations such as filtering,image enhancement and normalization were performed on the data sets,and three filtering processing methods were used for comparison,and the results showed that the median The results show that the median filtering process is more effective.（2）For the process of convolutional neural network used for artifact recognition and the common training methods,several techniques to prevent the model from overfitting are summarized;the structural characteristics of AlexNet and LeNet-5 network models and their advantages and disadvantages for use in artifact recognition are studied,and improvements suitable for artifact recognition in this paper are made for their disadvantages;in order to understand the role of convolutional kernel and the functions of each layer of CNN The feature map of the image output after convolution is analyzed by visualization operation.（3）To address the problems of slow training speed,fluctuating loss and acc curves,low accuracy in the validation set,and some degree of overfitting in the training process of AlexNet,we analyzed the causes and made several improvements to the network,reducing the network parameters by reducing the input size of the image,the number of network layers,and replacing the local response normalization layer with the current mainstream batch normalization layer.After the improvement,the recognition effect is significantly improved;for the problems of simple structure and low recognition rate of LeNet-5,the input image size and network depth are increased,the original activation function is replaced by ReLU,and the large convolutional kernel is replaced by multiple small convolutional kernels,and the results show that the overfitting phenomenon can be better prevented.（4）The experimental comparison proves that the improved network model can better recognize the artifacts in this paper,and the accuracy and recognition speed are greatly improved and meet the expectation,and the storage space occupied is smaller,the generalization performance is good,and the performance,recognition speed and accuracy of the model are not only better than the original network,but also better than the traditional machine learning algorithm.