Convolution Neural Network For Contented Based Image Retrieval

With the rapid development of mobile Internet technology,the number of pictures has exploded and content based image retrieval has become a research hotspot.The features of image used in traditional methods are manual designed,which are easily affected by noise.And convolution neural network(CNN)can complete the task of image classification automatically with the features extracted from massive images,which brings higher recognition rate and wider applicability.CNN simulates the human visual system to provide hierarchical abstraction for image to generate classification results.CNN makes use of local receptive fields,shared weights and spatial sampling,making significantly reduced training parameters of the network compared with traditional neural network and bringing a certain invariance against with translation,rotation and distortion.Nowadays,CNN has widely applied in image retrieval,which has a excellent performance compared to the traditional image classification methods.Therefore,the study of applying CNN to content based image retrieval has important significance.This paper is mainly focus on practical application and network improvement,the main tasks are as follows:(1)For the problem of CNN model design and network parameters setting,contrast experiments are setting in MNIST and CIFAR-10 database which adjust the model with different kernel number and size?pooling method and activation function.The experiments show that the more the number of kernel?the smaller the kernel size,the better performance the model has,and when the activation function using Relu?the first pooling layer using max pool,the accuracy is better.(2)For the classification of antique image,an image preprocessing method is proposed to solve the problem of target deformation which is caused by the different image sizes;We also find that when target is separated from background as the network input,the scale of the network will be much smaller and the accuracy will be better;There are some multiple target image in the antique image database,experiment about multiply target image is conducted which shows that CNN has a good ability to identify these images;Finally,a method based on the CNN and HOG-SVM is used for solving the unbalance samples problem,which has a better accuracy than the method of just using CNN.(3)In CNN pooling layer,there are advantages and disadvantages in different pooling methods.So this paper proposes a novel model named parallel pooling CNN(PPCNN).The model is designed based on LeNet5 model,which extracts features by using max pooling and average pooling at the same time in the first pooling layer.Compared to LeNet5,PPCNN model are more robust and has stronger presentation ability;In addition,a network training method is proposed to automatically select valid samples,which can solve the problem that the training network can not converge when the training samples are noisy.