An Application Study On Target Recognition Based On An Improved SIFT Feature Matching Method

With the rapid development of the computer technology, the target recognition technology has been widely applied. In recent years, among many target recognition methods, the feature matching method has been becoming more and more attracted attention. The SIFT feature matching algorithm is applied to recognize target which background is more complicated.The target identification of visible light image is studied in this paper, which contains three parts as followed: the image processing, the feature point detection and the feature vectors matching.In the stage of image pre-processing, on the basis of the characteristics of the image, an adaptive equalization image enhanced algorithm which based on the correlation coefficient is applied in this paper. This algorithm not only enhances the useful information of image but also preserves color information and luminance information of the original image.In the time of feature extraction, according to the traditional SIFT algorithm extracted more feature points, which leads to prolong the recognition process time and increase the false match rate. Thus, an improved SIFT extracting feature point method is presented. Firstly, the frequency significant regional detection method is used to indirectly get the feature points to replace the application of the DOG operator detecting the extremum points. Thereby the number of feature points extracted is significantly reduced, and the effect of the obtained feature points as the same as the human eyes. Which could guarantee the target feature points extracted are representative. Secondly, on the basis of it, the detected salient regions apply ellipse fitting to make the feature descriptors have the affine invariance. Finally used the Gauss weighted calculation is exploited accumulate gradient. In order to reduce the effect of the far pixels exerted on the feature points. The experimental results show that, the algorithm proposed by this paper is apparently decrease the number of detected feature points and obviously improve the feature ratio of the target object which compared with the traditional algorithmIn the period of feature vector matching, in order to ensure the two vectors havesimilarities in direction and distance. The similarity measure that the maximum entropy combined with the cosine method to search the Kd-tree is proposed. After completing the distance similarity, the method of cosine is applied to fatherly define the similarity in the direction to reduce the mismatch. To verify the proposed algorithm, the natural scene target is regarded as the application background to identify simulation analysis in this paper. The simulation results show that the algorithm proposed by this paper has higher recognition rate, faster processing speed and real-time processing capability.