Image Classification Algorithm Based On Gabor Wavelet And Convolutional Neural Networks

Image classification technologies such as character recognition,face recognition,and biomedical image recognition are very important in many computer vision fields.Accurate and efficient image classification is one of the hot research topics in the field of pattern recognition.Traditional recognition methods based on high-dimensional Gabor features have poor recognition performance.Although deep learning methods have achieved excellent results through parameter learning,they still encounter high computation complexity,long training time,and poor robustness against rotations and scale variations.With the adequate analysis on Gabor wavelet and convolutional neural networks,this thesis studies image classification based on Gabor wavelet and convolutional neural networks.The main research work and innovations of this thesis are as follows:1)In order to reduce the impact of high-dimensional Gabor features on real-time recognition system and to improve the robustness against rotations and grayscale changes,this thesis proposes a face recognition method based on binary-coded Gabor features.Firstly,the Haar-Adaboost classifier is used for face detection and posture correction is performed based on 68 key points of human faces detected by Dlib.Next,different Gabor wavelets are used to extract the multi-directional and multi-scale Gabor features.To reduce the dimension of extracted feature vectors,Gabor features with different directions and the same scale are fused by binary coding,and then principal component analysis(PCA)and linear discriminant analysis(LDA)are used to reduce the dimension.Finally,the support vector machine(SVM)is used as a classifier,and the experiment is performed on the FERET face database,which proves the effectiveness of the proposed method.2)Compared with traditional feature descriptors in the field of pattern recognition,neural network lacks specific domain knowledge and can only extract features based on data learning,so it often corresponds to long training time and high model complexity.In order to improve the training efficiency and recognition performance of convolutional neural networks,this thesis proposes an image classification method based on enhanced Gabor features and convolutional neural networks.Firstly,Gabor wavelets with different parameters are used to extract Gabor features.Secondly,the multi-directional and multi-scale Gabor features are integrated with grayscale image to build enhanced Gabor features.Thirdly,the enhanced Gabor features are used as the input of neural networks.Additionally,the parameters of Gabor wavelet can also be updated through parameter learning to reduce the impact of Gabor wavelet parameters on recognition performance.Experimental results demonstrate that using enhanced Gabor features instead of images as the input of networks can provide efficient shallow features and improve image classification performance.3)In order to improve the robustness of convolutional neural networks to transformations,this thesis proposes an image classification method based on Gabor convolutional neural networks.Firstly,the Gabor feature extraction module,the parallel convolution module,and the transformation pooling module are designed to combine the excellent characteristics of Gabor wavelets and convolutional layer.Next,according to the application scenarios of image classification tasks,the Gabor convolutional layer is constructed by choosing and combining these modules.Finally,Gabor convolutional neural networks are constructed by replacing traditional convolutional layers with Gabor convolutional layers based on baseline networks.Experimental results on MNIST,MNIST-rot,MNIST-scale,SVHN,CIFAR,and LFW datasets show that the proposed algorithm can improve network recognition performance and enhance the robustness of convolutional neural networks to spatial transformations such as rotations and scale variations.