Image Object Features Extraction And Recognition

Visual perception is an important way for human understanding the world, because the human observes the whole object and the information of object is concurrently processed by the nervous system, it is easy to identify the object in images for human eyes. Up to now the computer processes image data pixel by pixel under the lack of guidance of the global information. The accuracy and efficiency of computer automatic recognition of objects in images are yet far less than the human vision. Although many new research results of the area of computer vision have been reported in recent years, there are still many theoretical or practical problems that need to explore.Recognition of two-dimensional images is an important task of computer vision research. The aim is to identify the object by the methods of images analysis. Firstly segment the object from images, then extract the features of the object, and finally identify the object based on the features. So the key issues of object identification are image segmentation, object feature extraction, and recognition algorithm design based on feature.In this thesis, we focus on the problems of fingerprint segmentation, object shape description, identification of striated marks images of guns. The main work of this dissertation is as follows:1. A new fingerprint segmentation method based on support vector machine (SVM) is proposed in this paper. The algorithm uses the features of the gray variance and contrast to train the SVM for fingerprint segmentation, and fits the segmentation result of the SVM to convex hull. The experimental results show the algorithm is robust , especially for training segmentation algorithm with a small quantity of samples. The correct segmentation rate is 95% using FVC2002db4.2. An algorithm to estimate the quality of a fingerprint image is presented in this paper. At first, the original image is divided into local image blocks (32×32), the quality of each local image is estimated by the corresponding spectral distribution. Then the global uniformity and continuity of the image are analyzed. Finally , an image quality score is given by considering the local and global analysis results. Experimental results validate the proposed algorithm.3. A novel feature of relative direction of the object edge has been proposed in this thesis. The shape of the object is extracted and presented using with feature graph with key vertexs and the relative directon of edges linked to each vertex. Based on the feature graph, the algorithm of recognition is translation, rotation and scale invariant. Experiment results show that the characteristics is effective and fast, the average recognition rate of three gesture types is 98% and the average time of each image identification is less than 0.45ms.4. A new method of identification of striated marks images of guns presented in this paper. The automatical identification of striated marks images of non-standard machine gun of the image has not yet been reported successfully. The recognition of these marks is by professionals with the microscope and its efficiency is very low. The proposed algorithm recognize a non-standard rifle bullet marks image using SIFT features images. Experiment results show that the methos is effective and practical.