| 英文摘要 | With the extensive propagation of the Internet, the methods of human production and life have been undergoing profound changes. By means of the Internet, people can work, study, communicate with each other, and enjoy entertainments. Internet has been providing many conveniences for the human, meanwhile, lots of harmful contents have been emerging on the web, and overflowing seriously. These harmful contents have greatly injured human normal development, especially teenager's physical and mental health. Consequently, it's of great meaning to detect and recognize them. Generally, there are many kinds of harmful contents, such as pornographic information, violence information, horror information, drug information, reactionary information, etc. At present, little work has been done for the detection and recognition of drug information, however, their harm are comparable with, or even worse than those of other kinds of harmful contents. For this purpose, taking full advantage of object recognition and information fusion, we have conducted the classification of drug images and web pages on the Internet. The main contributions are as follows： 1. Drug-taking instruments recognition based on holistic features. Drug-taking instruments have apparent shape features, correspondingly, can be recognized by them. After comparing some shape descriptors， PHOG is used as the shape descriptor of drug-taking instruments. Taking PHOG as input, SVM can recognize these drug-taking instruments satisfactorily. The discriminabilities of four voting schemes in the generation of histogram of orientation are also compared. Experimental results on the PASCAL2011 dataset and the instruments dataset demonstrate that, only counting the number of pixels that belong to each bin can achieve comparable or even better performance. 2. Cannabis recognition based on local features. As a kind of plant, cannabis can present diverse appearance. Consequently, they should be recognized by local features. We compare the discriminabilities of two kinds of local features(LBP and SIFT), and find that using SIFT for cannabis recognition can achieve better performance. Five coding schemes of the BOW model are evaluated on a more challenging dataset, and the result is that, for cannabis images, the hard coding scheme achieves the best performance. 3. Saliency driven nonlinear diffusion filtering and its application in cannabis recognition and other objects recognition. Saliency image can present... |
修改评论