Research Of Improved Convolutional Neural Network Model And Its Application

Convolutional neural networks belong to deep learning architecture, which use local connections, shared weights and sub-sampling operations, so that the number of parameters need to be trained is reduced, and convolutional neural networks with more layers still have good performance. In addition, convolution neural networks also have shift invariance, scale invariance and many other advantages. So far, convolutional neural networks have been widely used and have achieved great success in many areas, however, the development of convolutional neural networks still have the following problems:Studies have shown that the activation functions have great influence on network performance, however, the selection of activation functions is very difficult. Currently, the suitable activation functions are mainly selected from a number of artificial activation functions by experience or experiments. If the function is to be selected by experience, on one hand, the experience may not be accurate, on the other hand, maybe there is no experience. If they are selected by experiments, since other parameters of training algorithms need to be selected, which require using cross-validation to be determined, the time to determine the best function and parameters increases significantly. To experiment, this brings a lot of inconvenience. For the difficult problem of activation function selection, this paper presents sparse maxout convolutional neural network based on maxout to solve the difficult problem of activation function selection and improve the performance of maxout convolution neural network.Generally, the learning of convolutional neural network would adopt a supervised learning or unsupervised learning.The unsupervised learning is cumbersome, time consuming, and the features of the training is independent of the specific tasks. And the supervised learning exists some problems, such as diffusion gradient. To solve the problems, this paper has a supervised learning and unsupervised learning algorithm combined and proposes a type of convolutional neural network based on k-means, which is then applied to traffic sign recognition later.The main work of this paper includes three aspects:1, for the difficult problem of activation function selection, this paper presents sparse maxout convolutional neural network, tested on the MNIST and CIFAR-10 data sets. The results of the test show that the sparse maxout convolutional neural network is more accurate than the maxout convolutional neural network, and the stability is better than the latter.2, for the learning problems of convolutional neural network, a convolutional neural network based on k-means is proposed, and tested on the MNIST and CIFAR-10 data sets. Experimental results show that in the case of comparable size, the accuracy of convolutional neural network based on k-means on the test set is higher than that of other algorithms.3, for the traffic sign recognition problem, combined with convolutional neural network based on the k-means,a specific algorithm is proposed to solve the traffic sign recognition problem. First, the images containing traffic signs need a preprocessing, including dimensions scaling to the uniform size, converting to gray scale and contrast enhancement. The next step is using convolutional neural networks based on k-menas to extract features. The last step is using SVM to classify. The algorithm is tested on German Traffic Sign Recognition Benchmarks. The test results show that the algorithm has a higher accuracy on the test set, and the size of the network is superior to other algorithms, that is to say, it has higher accuracy and lower time complexity.