Color Image Classification Method-Based On Feature Learning

With the rapid development of digital media technology in modern society,images provide great convenience for our life,which has also led to the increasing scale of image data that needs to be processed.How to design a feature descriptor that express image content more efficiently has attracted the attention of many scholars in the field of computer vision and pattern recognition.Local Binary Pattern(LBP)is a texture feature descriptor based on grayscale images.This method is favored by a large number of scholars because of its simple in principle,easy to understand,fast in calculation,and strong texture description ability.The color information of the image is ignored due to the LBP is limited to grayscale images.In order to further improve the performance of LBP,more and more researchers try to add color information to LBP.This thesis proposes a color LBP,which can describe color images efficiently when the feature dimension is low,and has a high noise robustness as well as higher illumination variation.On the other hand,this thesis also proposes a color LBP with rotation invariance.Specific research is as follows:In this thesis,we propose a Learned Color-Related Local Binary Pattern(cLBP)method for color image feature extraction.The method first calculates the Relative Similarity Space(RSS)by the RGB color space,the RSS and traditional RGB color spaces are combined to fully represent the color information.On this basis,the traditional LBP operation is performed on the six channels to obtain the corresponding LBP feature maps respectively.This is followed by a decoding step to capture the joint color information of those LBP feature maps,which corresponds to each color channel.Finally,a color-related pattern learning strategy is applied on the decoded LBP to improve the recognition accuracy and efficiency.Theoretical analysis shows that,compared with RGB color space,RSS can provide more discriminant information,and has higher robustness to noise and illumination variation.Experimental results on four groups of public color image databases show that cLBP is better than most existing color LBP methods in terms of number of features dimension,accuracy under noise-free,noisy and illumination variation conditions.On the basis of cLBP,this thesis solves the problem that the descriptor is sensitive to rotation,and proposes a rotation-invariant color descriptor,Learned Rotation Invariant Color-related Local Binary Pattern(cr LBP).With the characteristics of the color image,the cr LBP rotates the LBP binary string of each color channel synchronously to extract the feature vectors with rotation invariance.This method can avoid performance degradation of the rotation-invariant LBP,and the performance of the method is verified on 3 color texture datasets.