Research On Image Retrieval And Action Recognition Based On BoF

With the rapid development of the Internet, the number of the images on the Internet increases year by year. At present, there are billions of web images. How to quickly find out the image, which the users want to get, in such big scale of the image set has become an important problem in computer vision field. In order to solve the problem, the content-based image retrieval research is put forward and has received the widespread attention. And many new algorithms were proposed every year. This paper study the basic method of content-based image retrieval and the sub problem of image retrieval:action recognition in still image. At present, the most favorite way is to use collection of local features to describe the image. And the Bag-of-Feature (BoF) model can be used to turn the local features to the unified vector. All the researches of this paper are based on the extended algorithm of BoF model.The fusion feature for representing content of images has been investigated in order to optimize content-based image retrieval method. Firstly, Rootsift-based BoF is extracted, which capture shape and edge information. Then, Bag-of-Color (BoC) based on HSV is adopted, which capture color information. Lastly, BoC-BoF algorithm which integrate BoC vectors and BoF vectors by Gaussian external normalization algorithm is proposed. BoC-BoF algorithm effectively realizes the integration of global features and local features. The obtained impressive results show that this algorithm is more effective than other methods in two datasets of this paper.In order to optimize the content-based image retrieval method, this paper also studies its sub problem:action recognition in still image. The reason of researching the problem is that some top semantic retrieval results need to be improved. Though to combine the classifiers build by the key words of actions with the traditional retrieval model, people will get better retrieval result which in line with the high-level semantics. A novel action recognition method based on general multiple kernel learning is proposed in this paper. Firstly, HOG (Histogram of Oriented Gradients) based on edge of image and SIFT (Scale Invariant Feature Transform) based on dense sampling are extracted. Furthermore, spatial pyramid model is considered to obtain coarse spatial information. Then, the kernel matrix of each level in spatial model is computed by histogram intersection kernel function. With general multiple kernel learning, the weights of kernel matrixes are solved and the optimal kernel matrix is achieved by the linear combination of kernel matrixes. Lastly, action recognition is realized by the decision function. The obtained impressive result shows that the algorithm in this paper is more effective than some common methods in Willow-actions dataset.