Natural Image Classification Method Based On The Deep Learning Research

In recent years,with the development of scientific computer network and artificial intelligence field,the amount of graphic image data has been increasing day by day.How to extract visual information from a mass of natural images has become an important research subject in the field of machine learning.The classification of natural image data is one of the important studies to obtain natural image information.Convolution neural networks are an important application of deep learning in image processing.It is superior to other machine learning algorithms such as SVM,etc.It has the advantage of being able to convolve the image pixels directly,extract the image features from the image pixels,and be able to use the massive image data to train the network parameters sufficiently.Based on deep learning,this paper studies the natural image classification method.The main work and innovation are depicted as follows:1)Based on the tensorflow,a shallow convolution neural network for identifying images is designed.And the performance of the network is compared with single GPU and multi-GPU training,Thespeed of multi-GPU training reduced by 25 minutes.The intent of the work is to establish a better network structure for training and evaluation,and pave the way for creating a more complex network model.2)In this paper,the network structure and parameters of the convolution neural network are improved and optimized respectively.The experimental results show that the optimization of batch value,dropout,momentum and data set can improve the recognition rate of deep convolution neural network model.Therefore,it is very important to improve the network layer and optimize the training parameters to achieve the better training efficiency and the best classification effect.3)Based on the tensorflow and GPU to accelerate training the convolution of neural network structure,in this process it can improve the design and parameter optimization.First,a deep convolution neural network with a 9-layer structure is designed.Secondly,using cifar-10 and cifar-100 complex image database train and test the network structure,and in this process to optimize the parameters of the network structure.The results show that theclassification accuracy of the network structure is improved by 9.26% and 3.55% respectively compared with the previous network model(Conv-KN).The classification effect of deep convolution neural network under tensorflow frame platform is obviously better than other platforms,And the training time has also been greatly improved.