The Research On An Improved SIFT Feature Matching Algorithms Based On Spatial Structure And Color Information

The descriptors, which are extracted from the image by SIFT feature matching algorithm, are totally rotation-invariant and scale-invariant by detecting and describing the local feature of the image. So it is widely applied in the field of image processing. But there are also some drawbacks. Firstly, the descriptor are generated via SIFT algorithm in consideration of local features in the image, while the global spatial structure feature is not taken into consideration. Secondly, SIFT feature matching algorithm generates the descriptor in gray space without considering the color information of the image, which makes the algorithm not sensitive to the change of the image color. To solve the above drawbacks, we use singular value decomposition algorithm(SVD), the probability relaxation iterative algorithm(PRI) and comprehensive color image normalization algorithm(CCIN) to improve it. The paper has mainly completed the following work: 1. Modify iteratively the initial matching results of SIFT algorithm using the support for the initial matching results of SIFT from SVD matching results in space to improve its lacks of spatial structure information.Compared with the SIFT matching results by using partial least squares algorithm(PLS).The proposed algorithm has a higher matching accuracy without parameter constraints. 2. The color image is sensitive to illumination variation. This paper use comprehensive color image normalization algorithm to process color image from illumination variation. Then generate SIFT color descriptor with full color invariant to make up the missing of color information of the original algorithm. A large amount of experiments show that the SIFT color descriptor extracted by proposed algorithm are still totally rotation-invariant and scale-invariant. Compared with other SIFT algorithm combined with color information, the proposed algorithm is totally color-invariant. After reducing the dimensionality of the color descriptor by PCA algorithm and removing the mismatch points by partial least squares algorithm, the matching accuracy can achieve more than 90%. 3. Improved SIFT matching algorithm with spatial structure and color information. The adjacency matrix extracted from SIFT color feature descriptors is directly used to execute SVD matching.The global and local points are seen as matching points. The experimental results show that the proposed algorithm has a higher matching accuracy with no requirements to remove mismatching points.