Study On Image Similarity Measure And Its Application Based On SIFT Feature

Based on image similarity measure, the degree of similarity between two images is able to be assessed by comparing the features of image, such as color, texture, structure, semantics, et al. It is widely used in fields of pattern recognition, image retrieval and classification, image quality assessment and so on. In this thesis, using Scale Invariant Feature Transform (SIFT), the method of image retargeting and3D model retrieval are researched.SIFT is a type of local feature of image. It is invariant to image scale and rotation, and is shown to provide robust matching across a substantial range of affine distortion, change in3D viewpoint, addition of noise, and change in illumination. For assessing the similarity between two images with different sizes, two manners are employed. One manner uses the average of sum of the minimum Euclidean distance between matched SIFT features from two images; the other one utilizes the ratio of matched SIFT pairs number across two images relative to the number of SIFT features of original image. For the scenario of the same size images, from which lots of SIFT features are extracted, k-means algorithm is used for clustering and quantifying the features to a simplified vector, and then Kullback-Leibler divergence is used for assessing the similarity between the two images using the simplified vector.On the basis of the SIFT-based image similarity measure, an image retargeting algorithm combining Seam Carving with Scaling is studied. An image is resized using Seam Carving first. SIFT features are extracted from the original image and the retargeted one, respectively. And then, the SIFT-based similarity distance between the original image and its retargeted one is calculated. Before the salient object and content are damaged obviously, Seam Carving is stopped and Scaling is used for residual task. Experiments show that the algorithm in this thesis is able to avoid the damage and distortion of image content and preserve both the local structure and the global visual effect of the image graciously.Using a SIFT-based image similarity measure, a3D model retrieval algorithm is researched. First of all, a3D model is normalized in location and scale. By projecting the normalized3D model from multiple viewpoints, omnidirectional2D depth images are obtained and from which SIFT features are extracted. All the SIFT features associated with a model are clustered to generate a simplified vector as the feature vector of a3D model using Individual-k-means and Whole-k-mean respectively. By similarity measure of Kullback-Leibler divergence, an assessment of similarity between two3D models is achieved. 3D model retrieval is implemented by ranking images with the similarity values. Experiments reveal that the algorithm in this thesis is satisfactory to articulated shapes, rigid shapes and many other3D models.