Constrction Of Lightweight Convolutional Network For Image Classification

Image classification is a technique that uses a computer to find the essential image features and classify unknown images into different semantic categories.Image classification has been successfully applied in many high-level computer vision tasks,such as traffic scene analysis,medical image retrieval,face recognition and so on.As is well-known,images contain external disturbances including illumination,translation,rotation,scale and nonlinear deformation,etc.,so that the intra-class difference and the inter-class similarity are both large,which will greatly increases the difficulty of image classification.At present,the image classification algorithms based on Convolutional Neural Networks(CNNs)have attracted extensive attentions by their robustness feature extraction abilities.However,these good classification accuracy rates are resulting from mass of labeled data,such a process will limits its application in cases where the samples are scarce or difficult to label.In view of the difficulties faced by image classification and the shortcomings of existing classification algorithms,this paper mainly studies the following two aspects:(1)Constructing a Lightweight scattering convolution network based on Dual-Tree Complex Wavelet TransformIn order to solve the difficulties of image classification and obtain a comparable classification results to CNN with few training samples,this study constructs a lightweight scattering convolution network based on Dual-Tree Complex Wavelet Transform(DTCWT)for classification.Firstly,a DTCWT filter bank with multi-scale and multi-directional characteristics is used as predefined convolution kernels,which can avoid complex training processes meanwhile extracting rich information of images.In addition,the compact support characteristics of wavelets will introduce nonlinear deformation robustness to the network;Secondly,in order to obtain translation invariance,a nonlinear module operation is performed on the complex coefficients generated from the convolutional layer,and a log transformation is used to remove the influence of the outliers and illumination;Thirdly,extracting amplitude weighted relative phase histograms to maintain the rotation invariance;Finally,the features after Orthogonal Least Squares(OLS)process are fed into Support Vector Machine(SVM)for classification.The experimental results verify that the scattering network can obtain better performance with fewer samples,which provides a possibility for effective classification in data scarcity occasion.(2)Constructing a lightweight convolutional network Gabor-DCTnet based on Gabor and DCT filtersBased on the above research,in order to extract more abundant image features to improve the effects in face recognition and texture classification,this paper constructs a new lightweight convolution network Gabor-DCTnet.In the convolutional layer,the inter-stage synthesized Gabor and DCT basis are cross-convolved to obtain a complex filter bank FBGabor_DCT,then used it as predefined convolution kernels to extract rich features,including textures,edges and contours etc.Then cascading nonlinear operation,in this layer the convolutional coefficients are binarized to obtain illumination robustness,and the binarized features are hash coded for feature fusion and dimensionality reduction.After that,block-wise histograms are performed to introduce rotation invariance.Average pooling are used to average the histogram features of real and imaginary parts for obtaining more discriminative features and dimensionality reduction.Finally,Nearest Neighbor classifier is employed to classify the unknown images.Experiment results show that,Gabor-DCTnet is do better than other lightweight network on face datasets including FERET_?,FERET_?,AR and texture dataset including KTH_TIPS and CUReT,which demonstrates that the learning-free Gabor-DCTnet is suitable for both face recognition and texture classification tasks.