Research On Image Retrieval Based On AdaBoost And SVM

With the popularity of Web technology and the development of information technology, multimedia information, such as images, sounds and video, has quickly become the mainstream way of information dissemination and exchange. A large number of image and video information stored in the large-scale image database, which makes image database becomes large and complex. How to select semantics and visual features of the image from the desired image is currently facing a pressing problem. And Traditional keyword-based query methods are far from meeting the modern image information retrieval. The traditional database technology is facing the challenges of organization, expression, storage, management, query and retrieval of data information of these images. Therefore, the combination of image processing, pattern recognition, computer vision detection technology and database technology to establish efficient image retrieval mechanisms and develop good performance of the retrieval tools has been becoming an urgent which is need to address the problem. Content-based image retrieval technology has become the key to solving this problem.The target detection of content-based is a hotspot research in the field of pattern recognition. It not only has a good theoretical significance, but also has a broad application prospects in many areas like health care, transportation, and many other fields of investigation. With the deepening of the research and the development of identification technology, some technology based on target detection has matured and a number of practical applications have appeared, Such as medical image retrieval system, trademark retrieval system, face detection and recognition, text or handwriting recognition, fingerprint detection, intelligent transportation systems and so on.For the problems of the AdaBoost cascade algorithm with the increased difficulty of learning likely to cause over fitting a result of the decline in the efficiency of the classifier, poor stability and the problem of SVM facing large-scale samples needs a the long time for training, playing the unique advantages of SVM in the small scaled, nonlinear, high dimensionality samples. We improved the standard AdaBoost cascade classifier training algorithm and proposed an AdaBoost-SVM cascading classification algorithm in this paper. In the beginning of the AdaBoost cascade classifier algorithm, we set the maximum number of weak classifiers for each level, when the number of weak classifiers reached the maximum number in a certain level but the classifiers cannot achieve the goals of the false alarm rate, the AdaBoost classifier are replaced by SVM classifier, then the SVM classifier can just train the samples that AdaBoost classifier has selected and the training time will be shapely reduced. So AdaBoost-SVM cascade classification algorithm can not only get the higher classification accuracy and low false alarm rate, but also ensure that the training speed and detection rate.In experiment, we choose two different types of images or face recognition and pedestrian detection experiments. And the candidate areas with a window of fixed size were captured for integral image and Haar-like rectangle feature extracting in the experiment, and then two classes classification was performed by using the proposed cascade AdaBoost-SVM classifier. The experimental result shows that the cascade classifier proposed by us can get better performance than cascade AdaBoost classifier in accuracy and the false alarm rate, compared with SVM classifier, the training time is greatly reduced.