Technology Research, Based On Local Features For Image Object Recognition

Image target recognition is one of the most important topics currently in the domain of image processing and pattern recognition. There are many limits in practical applications when global features are used in the present system. So the excellent performance of the local features provides an effective way for solving the recognizing target in complicated images. In view of the deficiencies of the research results at present, this dissertation focuses on the problems of using local features for target recognition, including local features extraction, target matching, target classification and target recognition with changes in 3D viewpoint.Target feature extraction, as a key technology of target recognition, has a profound influence on the eventual performance. Because most local features should be invariant to luminance, translation, rotation and scale change, the extraction of local features depends on the specific problems and the knowledge in the relevant fields. Based on the discussions of the most fashionable local features, some improved region detection algorithms and descriptors are proposed.This paper presents an algorithm of image mosaic which registers images according to multiresolution analysis and local features. Then the problem of image mosaic is changed into another problem that the coordinate transformation between the pixels by affine transformation. An image retrieval algorithm based on local features is also proposed. The local features are optimized into"prototypes"by the technique of relevance feedback. Then the method enhances the accuracy and efficiency of features matching by comparing the distance of the closest neighbor to that of the second-closest neighbor.The local information described in local features offers a clear description of the image contents on several semantic levels. To the deficiencies of the target classification based on the local features, we present a method for target classification on the basis of succeed experience about the vector space model used in text categorization. We also optimize the feature space by the information theory and related techniques. Experimental results on a standard database show that this algorithm is effective and robust.Visual images of one object will be different when the 3D viewpoint has been changed. We use a relative 2D aspect set of a 3D model to reflect the different appearances of the object in various views. Then an algorithm for pattern matching based on the Hausdorff distance is proposed. Meanwhile, a method for extracting corner feature from images is proposed. This feature is invariant to translation, rotation, scale change and is shown robust to addition of noise. We present a system to recognize the object with changes in 3D viewpoint using this feature and BP network. The performance on the obtained experimental results demonstrated that this method is more effective than the other three ones.