Research Of Image Classification Technology Based On Convolutional Neural Network

With the rapid development of mobile Internet technology, along with the extensive use of the rise of micro-channel, microblogging, mobile applications and other mobile media, as well as smart phones, tablet computers and digital cameras and other mobile devices with 3G, 4G, Wifi and other high-speed wireless network popularity, users can quickly and easily browse or upload images. However, there are a large number of unmarked difficult in real life image search and processing. But the traditional method of image classification results, to search people’s lives or bring a lot of inconvenience, especially in complex environments to identify natural images of people eager to image recognition can have a new breakthrough.Convolution neural network is popular images by category method(one deep learning methods), which draws on learning the principles of image processing neurons. Appears at the current deluge of image data, with more image samples with the hardware upgrade, just large-scale training brought great opportunities for neural networks. In this paper, based on the traditional handwritten digit classification framework LeNet-5 to improve, through classification task MNIST database and CIFAR-10 database convolution factors affect the performance of neural networks in; and redesigned based on neural network convolution a deep convolution network, to achieve good results in the Tiny ImageNet database.This innovation follows:1. For network model applied to image classification convolution design and how to choose the hierarchy optimization and other issues, according to the architecture we LeNet-5, and on its basis the introduction of a five-layer structure with a convolution neural network and application In MNIST database and recognition tasks CIFAR-10 database, the hierarchy, the activation function improved by adjusting the convolution neural network, descent algorithm, data enhancement, the pool of selection and feature maps to compare the number of experiments, the study found convolution neural network using pooled size of 3 * 3 and more cores(64 or more) and a small receptive fields(2 * 2), increase the level of structure, Relu activation function, driving amount of gradient descent algorithm and enhanced data After collection, under certain experimental conditions, the image classification results in MNIST database dropped 1.08 percent error rate, in CIFAR-10 database dropped 28.12 percent error rate.2. For Tiny ImageNet database classification of natural images, through the analysis of factors convolution neural network algorithm summary parameters and optimize the use of the law, and the design of a network layer depth level of 16(not including the pooling layer) depth volume area network, hierarchy 13 convolution layer and three-layer fully connected layer use. Compare Zeiler model(32.27% correct rate of Top-1) in Tiny ImageNet database photogenic, the proposed model has achieved 54.38% of Top-1 accuracy.