Study on extraction algorithm based on SIFT feature points
|School||Shaanxi Normal University|
|Course||Computer System Architecture|
|Keywords||SIFT geometric blur descriptor diffusion distance feature match|
In the age of the Computer intelligent, image research in the field of military, scientific research or industrial production has become more and more frequent. Image technology is changing the human way of life, and the impact on our daily lives. The consistency of the key points of the image plays an important role in the application of computer vision. The consistency of the key points of the image for image type recognition technology is one of the important issues in the field of computer vision.On the basis of in-depth analysis of the existing image feature extraction methods, the paper focuses on SIFT(Scale Invariant Feature Transform) feature of the type of image recognition application. In order to improve the effect of matching the type of the object in the image, the paper proposes a new feature descriptor, and use a new distance measure to determine the feature vector similarity. The main innovations of the paper are as follows:For the requirements of the type of the object in the image, paper analyses and points out that the existing image feature descriptor technology is not a good description of the characteristics of the object of the same type, and combines the theory and practice of the feature descriptor proposed using geometric blur descriptor to represent the local features of an image. Through the four separately gradient direction of the original image, feature descriptor is obtained. This feature descriptor effectively enhances the description of the local features of objects in an image.Traditional SIFT algorithm uses Euclidean distance to measure the feature operator between the two images. Euclidean distance does not correctly reflecting the mapping from the high-dimensional space to a low-dimensional structure, easily leads to matching errors. To overcome this drawback, the diffusion distance replaces Euclidean distance as a measure of the standard of distance. The algorithm finds two images feature vectors nearest neighbor and sub-nearest neighbor using K-d tree, and then using a random sample consensus algorithm to eliminate errors from the candidate matching match.By programming algorithm, the experimental results show that the algorithm can better reflect the related parts in the two images and match better than the original SIFT algorithm.