Image Classification Based On Neural Network

With the development of multi-media and Internet technology,the digital image has become an important information transmission medium.With the explosive growth of image data,how to analyze image content automatically has become an urgent problem,and therefore the image classification techniques emerged.According to the level of abstraction of features,image classification methods can be divided into three types,namely,low-level feature based methods,mid-level feature based methods and high-level feature based methods.This paper focuses on artificial neural network and studies image classification algorithms based on mid-level and high-level feature,respectively.From the perspective of mid-level feature,we propose a new feature representation method called Bag of Autoencoder Words(BoAEW).Bag of Visual Words(BoVW)is a classical image classification framework.Traditional BoVW methods usually use SIFT descriptor as low-level feature,which may prematurely lose useful information since SIFT is hand-crafted feature.Instead,learning low-level feature directly from image pixels is a more effective approach.With this idea,this paper presents a novel BoAEW visual feature by incorporating Autoencoder neural network into the BoVW framework.Instead of the traditional hand-crafted local features(such as SIFT),BoAEW employs Autoencoder to generate more adaptable local features.In this way,BoAEW not only broadens the scope of BoVW from the pre-defined feature to the unsupervised learning feature,but also makes the Autoencoder more practical as well.Extensive experiments on both face recognition and scene classification datasets demonstrate the effectiveness of BoAEW.From the perspective of high-level feature,this paper comprehensively studies texture classification with convolutional neural network(CNN).Recently,CNN has achieved great breakthroughs in many computer vision tasks.However,its application in texture classification has not previously been thoroughly researched.To this end,this paper carries out a systemic research on image texture classification.Specifically,CNN is used to extract preliminary image feature,and subsequent PCA operation can reduce its dimensionality to obtain final texture feature which is fed into a SVM classifier for prediction.The effectiveness and drawbacks of CNN are demonstrated by comprehensive experiments and analysis on four benchmark datasets.The results show that CNN is a favorable texture feature representation and achieves quite good performance in most image texture datasets,however,it performs worse in datasets with image noise and rotation.Thus,it is necessary to enhance the abilities of noise tolerance and rotation invariance of CNN and to construct a large diverse texture dataset to guarantee its best performance in image texture classification.