Research On Image Classification Technology Based On Deep Learning

With the rapid development of the Internet and digital media technologies,massive image data are emerging constantly.If these image data can be effectively classified and managed,the efficiency of querying image information can be improved,it's also beneficial to the following issues like tracking,segmentation,scene understanding and other complex visual problem.Deep learning is a method of learning features,which uses a multilevel network structure to learn the high level of expression of data by itself.Compared with the traditional method of manual design of features,the features learned by deep learning have stronger classification ability in image classification.This paper focuses on the research of image classification technology based on deep learning.The main work of this paper includes the following four points:(1)Aiming at the problem of slow convergence and low accuracy when classifying real world pictures for image classification models based on the Visual Attention,a low level convolution module is introduced.On the one hand,it's used for model initialization,which can make the model focus on the important regions in the early steps,and can improve the convergence rate and accuracy of the model.(2)In order to solve the problem of large computational cost in the image classification model based on the convolutional neural network,a shared convolutional structure is designed.The structure can share a part of the low level convolution operations,which reduces the computational complexity of the model and make the convolution deeper.(3)For the problem that similar samples are easily misclassified,a hard example mining training method is designed,which improves the ability to distinguish difficult samples.(4)Based on the RAM(The Recurrent Attention Model)and DRAM models(Deep Recurrent Attention Model),the above three points are merged to propose a CRAM model.This paper compares CRAM,RAM,DRAM and Network in Network model with Cifarl 0 dataset.Experiments show that CRAM can effectively improve the accuracy of classifying real world images while increase a small amount of computation on RAM and DRAM,and CRAM can speed up the convergence.In addition,the accuracy of CRAM is close to the Network in Network which has the same depth of convolution neural network with CRAM but takes up more computing resources.