A Novel Non-loss Function Deep Convolutional Neural Network Based Image Feature Extraction Method

As the one of artificial intelligence technologies,image recognition has been widely used in many domains of society.Such as code of container recognition,face recognition,Pathological image recognition and so on.For an integrated image recognition system,the main procedures can be divided into three parts:image acquisition,image feature extraction,image recognition.Especially,image feature extraction is the most important part in whole system,which directly affect the accuracy of final recognition.In recent years,with the development of deep learning,more and more image recognition algorithms adopt deep learning methods for image feature extraction.Because the shared weights of filters can reduce the computation consumption and extract local features in images,deep Convolutional Neural Networks(CNNs)become the most popular method in image feature extraction.However,as the current research trend towards multi-layers,more complex structure than before,it is harder for us to train a deep network.Researchers not only need to adjust the more parameters in deep model than shallow model,but also integrate current training methods with the old methods,such as dropout,maxout and so on.Recently,with proposal of conception of simple deep learning,more and more researchers start to focus on this kind of non-loss function deep learning models.The most famous one is PCANet which is proposed by Prof.Ma Yi,the characteristic of non-loss function makes training deep models easily.However,PCANet adopts unsupervised learning method,Principal Component Analysis(PCA)to train convolutional filters,which makes the recognition accuracy is not very high.Although Professor.Ma Yi also proposed Linear Discriminant Analysis(LDA)based deep model,LDANet,However,because LDA cannot separate positive and negative samples distinguishingly,it makes no remarkable improvement in recognition accuracy in LDANet.In addition,the way of down-sampling of PCANet and LDANet,block-wise histogram,easily makes deep models occur over-fitting phenomenon,which affect training deep models.According to the above problems,we do the following two aspects in this paper:1)In this paper,we creatively introduced Marginal Fisher Analysis into training filters and proposed a PCANet based novel simple deep learning model:MFANet.As Marginal Fisher Analysis(MFA)is to find out an optimal set of eigenvectors which can converge the intra-class projection and separate the inter-class projection.We use the computed eigenvectors as the filters of MFANet.The proposed model is evaluated by three datasets:ICDAR 2003 dataset,PIE face dataset and Scene-15 dataset.In addition,we compared our model with other popular models and showed the recognition accuracy of ours are better than those.2)In this paper,through testing different projection function,we used the Guassian function to replace Heaviside function in basic PCANet.By this way,we resolve the rigid boundary of Heaviside function in coding features.Meanwhile,we propose a new down-sampling method called Block-wise Stochastic Histogram(BSH)to resolve over-fitting phenomenon of basic PCANet models.The experimental results show that comparing with other down-sampling method,the over-fitting phenomenon of deep models which used our proposed BSH are less than those which used traditional Block-wise Histogram and other down-sampling methods.