Image Local Feature Extraction Based On Particle Swarm Optimization

Image feature extraction is a fundamental step of image matching, object detection and object recognition. In the fields of image feature extraction, local feature is a hot topic, which many interesting works have been proposed. Typically, good features should express the information of the image local areas which is invariant to view angle changes, illumination changes, scaling, rotations etc. In this paper, we analyzed the state-of-art methods in image local methods. Moreover, based on traditional SIFT and ASIFT, the task of feature matching has been transformed to parameter optimization. As a result, we improved the efficiency of feature matching by using particle swarm optimization and robust estimation. Details of our works are given as follows.(1) Affine SIFT based on particle swarm optimization is proposed in this paper. ASIFT is an extend version of SIFT based on affine sampling. As for SIFT, ASIFT is invariant to rotation and scale changes. Moreover, ASIFT is fully affine invariant, since it uses affine sampling to simulate viewpoint changes. Although ASIFT is powerful, we found that ASIFT uses discrete sampling to generate affine samples. Note that the parameters in affine transformation is continuous, namely, the transformation generated via ASIFT doesn’t describe the relationship between reference and input images. Therefore, a PSO based algorithm, which aims at searching the best combination of affine parameters, has been proposed. The experimental results show that our method find more matches than traditional ASIFT, especially in multi-view images.(2) Improved Multi-GS multi-structure estimation method is presented. To eliminate mismatches in PSO-ASIFT and to evaluate the transformation model between the reference and the input image, we analyzed the least square, Hough transform, RANSAC and Multi-GS algorithm. Therefore, based on the concept of Multi-GS, reverse order has been introduced to evaluate the similarity value between feature points. As for Multi-GS, we use ordered residuals, which have been computed via randomly generated structures, and reverse order to evaluate the similarity. The advantage of our method is to avoid tuning the value of h, which is the length of sub-window. The experimental results conducted on feature matches of the PSO-ASIFT revealed that the proposed method improved the efficiency of Multi-GS.