Research On Image Local Invariant Feature And Its Application

Image processing is the core research scheme of computer vision. It mainly refers to image pro-processing, feature extraction and description, feature combination and selection. In recent years treated image data as a global feature combination is not suitable for the complex application any more. In Marr's computer vision theory, image is treated in a local approach, which is also supported by the result of research on the human vision simulation. Using local feature to recognize image content has become a popular approach, which can significantly improve and accelerate different kinds of image object detection method. Now image processing based on local feature has been widely used in image pre-processing, image retrieval, remote sensing image processing and pattern analysis. In this paper we focus our research on explaining and recognizing image in a local approach.The highlights and main contribution of the dissertation include:(1) Research on Scale-space Theory: Scale-space Theory is the direct theoretic basis of local feature extraction. Early Scale-space idea is mainly been used in an experimental way without any obvious physical meaning and mathematical foundation. In this dissertation, we begin our research with mathematical analysis of differential arithmetic operators. Through the modeling of automation of scale selection and research on 1~2 dimension signal scale selection progress in scale space, we explain the idea of local invariant feature detection and the source of feature invariability. At last, the thesis gives an explicit physical definition of scale normalization; indicates the use of second moment matrix and Hessian matrix in local feature extraction.(2) Research on local invariant feature detector. Through modeling three kinds of typical local image structure and analyzing their behavior in scale space, we get a generalized progress of local invariant feature detection. The dissertation also proposes a new way to make use of local consistency in feature point localization. At last, an empiristic parameters selection solution is proposed to avoiding error caused by discrete scale space.(3) A local invariant feature descriptor based on Gaussian differential operators is proposed. In the dissertation, we analyze the function of Gaussian differential and requirement for distance function selection. In order to provide features with rotation invariability, a direction adjustment algorithm is proposed to estimate local dominant direction. Experiments show the efficiency of our proposal.(4) A local invariant feature based on salient estimation of local second moment matrix is proposed. In recent years, global feature extraction is not suitable for complicated image application, with the development of pattern analysis and machine learning, local invariant features have been widely used as an input for pattern recognition algorithm. Compared to the old method to extract local feature in a fixed window, local invariant feature approach start a complete new way to extract image feature in form of combination of scale-space theory and vision salient. It has become one of the dominant approaches to content based object recognition. In our paper we propose a new local invariant feature based on local area feature saliency measurement. According to the analysis of local second moment matrix eigenvalues'properties, we get the measurement of feature saliency, and thus propose a new local invariant feature which give a better appearance in real scene image test.(5) Remote sensing image processing based on local invariant feature. Multi-target detection and ROI quick match technology is one of the most challenging research domains. Different from traditional pre-model object recognition approach, our paper proposes an innovative way to detect and descript multi-targets based on local invariant feature. we use feature matching algorithm to complete ROI quick matching, and area adjustment in the meantime. In the dissertation, we eliminate error parameters by utilizing feature's geometric constraint. The algorithm overcomes the shortcomings of dependency on geographic information database. In the target detection, we incorporate the idea of local invariant feature into the detection of ships in the harbor area. Through a new feature extraction operator with scale information, we extracted salient area in remote sensing image and utilize semantic feature to localized ships. Experiments results show the efficiency of our algorithm and its great future.