Research On Image Classification Based On Kernel Laplacian Sparse Coding

With the advent of digital information age and the development of computer multimedia technology,A variety of image data increased dramatically.Thousands of images are produced daily,Image classification is necessary.The goal of many computer vision tasks is to estimate classification properties from visual data,Such as the presence or absence of a specific object in the scene,as well as in the case of a self-driving vehicle,one of the key tasks is to accurately identify cars,traffic signs and pedestrians.So the challenge of image classification is great.Because of the extensive application of image classification,making it in the visual field of rapid development.Sparse coding proposed has been widely recognized and applied.The main research results of this paper are divided into the following aspects:(1)Aiming at the lack of correlation between image features in sparse coding methods,laplacian regularization is added to the sparse coding method to encode local features,similar features have similar sparse coding words,and the method is naturally deduced by relaxing the restrictive cardinality constraints of vector quantization.The features and dictionaries are then mapped into high-dimensional spaces and the kernel Laplacian sparse coding is chosen to derive the image representation.Firstly,compared with vector quantization coding,sparse coding can achieve lower reconstruction errors due to less restrictive constraints.Secondly,the sparse coding of the image is counted through the maximum pooling method,which preserves the salient features of the image.Thirdly,using sparse coding to express the image,meets the nature of image sparse characteristic.(2)In view of the lack of spatial information in visual vocabulary features,a method of averaging regions is proposed,and then the encoded feature vectors are fused using maximum fusion.Similar to the histogram construction of the space pyramid,this is done in the largest area averaged over the image construction.By averaging the images and then make maximizing pooling,the aggregated features are more robust to the local transforms than the average statistics in the histogram.The overall structure of this paper is based on Laplacian sparse coding algorithm.The final representation of the image is made up of features brought together from a variety of locations.(3)In terms of classification accuracy,the method of averaging Laplacian sparse coding based on SIFT descriptors is significantly better than linear space pyramid matching kernels on the histogram,even better than non-linear pyramid matching kernel.By using a single type of descriptor,do the experiment under different conditions.Linear kernel support vector machines with faster training and testing speed,compared with the non-linear,significantly reducing the demand for the kernel.