Research On Image Local Feature Matching Enhancement

As human's living behavior cannot be separated from the eyes' detection of the real world.With the rapid development of computer technology,researchers who based on the characteristics of human vision,developed a series of applications to recognize reality,which have been widely used.The computer can save all of the input images' information for subsequent matching applications,but with the arrival of the era of large data,simple image preservation can no longer satisfy with people's request.Image recognition technology is more dependent than before,such as the massive image recognition and retrieval technology,the researchers have proposed a series of image feature extraction and description techniques based on the sparse coding in biometrics.In the proposed image feature processing methods,there are still many parts needed to be improved.For example,some traditional methods only apply to the description of grayscale images,which lead to the lack of color information for the description of the image.Image have their corresponding expression at different scales in the whole space,Many traditional methods describe image just from the selected single-scale layer,which lead to the loss of the image feature description information.At the same time,the gray gradient of a single-scale extracted image lacks the expression of salient information.Although using the image feature extraction techniques to obtain sparse image feature improve the efficiency of information representation,however,there are still some redundancy in these features,which result in the decrease of image feature matching accuracy,the waste of storage space and the lower speed of recognition matching.In order to overcome the shortcomings of existing methods of image feature extraction and description,this paper improves the performance of image feature processing algorithms in two aspects including four methods.The main research results are listed as follows:1.A image local feature descriptor based on biological visual mechanism is proposed,which combines color and shape features.Compared with the previous descriptors,red-cyan cells(L,M,S cones)are used as a channel of color opponent.Then,combined with the new color feature extraction technique,a new color feature extraction technique which calculating the color direction and amplitude of three color opponent channels in two-dimensional space is proposed.(Color opponent channel is red-green,blue-yellow,red – green,seperately).The color direction weighting with pixels' amplitude is composed of the histograms that calculated from the feature points and the surrounding pixels' directions.Finally,we use the fusion method to combine the Four-Channel-Opponent-Color and the scale invariant feature histogram.Experiments show that the proposed method is robust to photometric and geometrical changes compared with other descriptors,especially in the case of illumination and image blur changes,where color contrast information is predominant.2.A feature descriptor with multi-dimensional entropy of local image is proposed.The traditional image local feature description method is based on gradient information of image,ignoring the other significant content of image.In this case,the descriptor proposed in this paper is based on local image gradient descriptor,and concatenating local entropy information of feature points via later fusion method linearly.Althrough this method efficiently increases the significant information's expression for feature gradient descriptors,but the main problem of this method is that the range of the entropy descriptors is variable,which will lead to mismatching.Thus this paper add exponentially normalization for the entropy feature,the mean average precision of the descriptor is used to determine the dimension of the entropy feature and the parameter of exponent in the normalization.Compared with the traditional description method,the descriptor with multi-dimensional local entropy feature proposed in this paper has strong robustness to the artificial scenes' changing.3.Two local entropy-based image feature descriptors are proposed in multi-scale spaces.Because the traditional feature description methods,such as SIFT,describes feature only on a single scale,which will lose some important information and affect the experimental results of image matching.According to the property of entropy,we use information entropy to estimate feature points and their surrounding information to obtain more key content.The new method consists of firstly calculating the multi-layers' SIFT orientation histograms and local entropy values of the feature points in difference of Gaussian space(DOG),then at each layer,computing the percentage of feature point's entropy value within all layers' entropy summation,and multiplied the value by correspondent descriptors,finally all the obtained descriptors of the same feature point are summarized to obtain a new local entropy-based feature descriptor.The second local entropy descriptor is obtained by using the Hellinger distance in the new local entropy-based feature descriptor.Through the comparative experiments,precision-recall curves,mean average precision and the number of correct matching of our proposed methods have great performance.4.A salient maximum entropy-based image local feature descriptor is proposed.Since the feature points extracted by local gradient descriptor still exit noise information,such as edge effect points.the descriptor proposed in this paper is to extract the feature points in the salient parts of the image,the optimal threshold of these salient parts are selected by graph-based visual saliency.Then the entropy space is established based on the characteristics of the feature points.The histogram of the maximum entropy corresponding to each dimension descriptor in the scale layer is extracted to form the final descriptor,and the subsequent comparison shows that adding salient image segmentation and maximum entropy estimation to local image feature descriptor improves mean average precision in image matching.