Research On Feature And Similarity Measurement Of Computer Vision

Computer vision focuses on making meaningful decision about objects and scenes from perceived images, thus it can simulate human vision perception. Human vision gets perception from images by abstracting and coding features then evaluating the similarity between the features and prior knowledge. Accordingly, feature design and similarity measurement attracted considerable research attention.The thesis is concerned about feature design and similarity measurement in computer vision, and their application in non-rigid shape point matching and image fusion quality evaluation. The main contribution are summarized as follows,The state-of-arts of features and similarity measurements which are proved to be effective efficient and universal are introduced.For shape context feature, its weakness is analyzed and the reason caused it is found. We design a new feature which is invariant for translation, scale and rotation. The physical theories of the new feature are explained detailed and the properties are proved mathematically. The shape point matching experiment proves the effectivity and efficiency of new feature in application.Earth mover's distance can make simple shape context feature invariant to rotation, but its computation cost is too large. A scheme to calculate the earth mover's distance for shape context is proposed. Its computational complexity is much less than the traditional scheme.Image fusion is an important part of information fusion. As more and more fusion algorithms have been proposed, quantitative evaluation of the performance of different schemes becomes a crucial task. Especially the objective evaluation metric is a significant application of computer vision.The current status of image fusion and image fusion performance evaluation research at home and abroad are detailed introduced. We analyze the vision model of structure similarity based evaluation metric, find it oversimplify the characteristic of vision perception and doesn't consist with monitor properties. For this problem, new luminance and contrast models are employed to design a new image fusion performance evaluation metric. The experiments prove that the new models consider more of the environment influence and the monitor properties, the metric based on them is better compliant with subjective evaluation. The universal un-structure similarity based image fusion performance evaluation metrics are introduced. a new metric which based on a likelihood function.For color cue then is presented. The similarity of color distributions are measured by Minkowski distance. We provide a theoretical analysis on how the quality metric responds to weighted averaging fusion algorithm, and employ slide window with local saliency weighted to overcome its weakness. Extensive experiments have demonstrated that the metric is more effective and consistent with subjective evaluations compared with other metricsThe similarity measurement is applied to digital image scrambling. We propose a digital image grey level scrambling algorithm based on the classical Vigenere coding theory, the properties of the algorithm are analyzed and its effectivity and efficiency are examined by histogram similarity.