Research On Middle Feature Representation Based Image Classification

With the rapid development of Internet technology, it accumulates many digital images and videos in the network, which bring new technical challenges. For this reason, image processings, like image classification, image retrieval, and object recognition, have been drawing more and more people’s attention. Image feature extraction, which is used to map an image to a corresponding feature space, is the basis of image processing. At present, many excellent algorithms have been proposed, which can be classified according to different semantic expressions. Many researchers pay more attention to the mid-level semantic feature representation method for its good performance. In this paper, we analyze and improve some typical low-level and mid-level feature representation methods. Experiments demonstrate that our method has a better performance than other related methods. The main works are as follows:Firstly, we propose a new mid-level feature method by combining spatial and semantic information of image based on the existing bag-of-words (BoW) model. As the BoW model assumes that visual words are independent of each other, it ignores the relation between the visual words. In our method, we combine spatial information and semantic information. On the one hand, we extract similar visual words by computing distribution divergence, and form one visual phrase which includes image semantic information. On the other hand, we extract some semantic visual phrases from all visual phrases to constitute the phrase dictionary. The new mid-level feature representation is generated by the two kinds of information. Image classification experiments have been conducted on UIUC-Sports8dataset and Scene-15dataset, and the results show that our method has achieved better classification accuracy.Secondly, we improve Local Binary Patterns (LBP):1) we propose a texture phrase method based on LBP,2) we improve a BoWL-based feature representation method by combining spatial and semantic information. The experiments show that our methods can get better performance for image classification. At last, LBP is applied to the actual network images for its simple computation and easy expression. Experiments confirm the effectiveness of LBP to express the network textual images.