Research On Semantic Model Based Image Classification Method

Image classification is a key technology of image processing, such as image retrieval, image recognition, object localization and so on, and it is also an important research in the field of computer vision. Currently, image classification based on semantic model mainly aims at modeling the middle semantic of images to improve the analysis capability of image spatial information and narrow the “semantic gap” between low-level features and high-level semantics. This thesis mainly researches on image classification based on semantic model, including image classification based on Bag of Visual Words(BoVW) and image classification based on Visual Language Model(VLM). The contributions are listed in three aspects as follows:(1) Aiming to the synonymity and polysemy of visual words, an image classification method based on adaptive soft assignment Bag of Visual Words is proposed. Firstly, when local features are mapped to visual words, fuzziness is analyzed, and different number of visual words is allocated to reduce the synonymity and polysemy of visual words; Then, the visual stopping words are detected via chi-square model and removed from the dictionary; Finally, the visual vocabulary histograms of images are constructed and input to classifiers to complete training and classification. Experimental results show that, this novel method can effectively reduce the fuzziness of local features and the synonymity and polysemy of visual words, and improve the performance of image classification based on Bag of Visual Words.(2) Aiming to the ignorance of long distance dependency of visual words in traditional 2-gram Visual Language Model, an image classification method using N-skip Visual Language Model is presented. Firstly, the long distance visual word pairs are trained, and the conditional probability is estimated; Then, the spatial information is expressed via fusing different weighted visual word pairs; Finally, the categories which the images belong to are determined via maximum likelihood estimation. Experimental results show that the new method introduces the long distance dependency to represent image content, which enhances the capability of expressing the spatial information of visual words and improves the performance of image classification based on Visual Language Model.(3) Aiming to the ignorance of image background noise in traditional Visual Language Model, an image classification method based on weighted-saliency-map Visual Language Model is put forward. Firstly, the salient regions and background regions are extracted via saliency detection algorithm based on visual attention; Then, the visual documents of images with salient labels are constructed, and the salient weights and conditional probability are estimated in the training phase; Finally, the images are classified with weighted-saliency-map Visual Language Model. Experimental results show that the novel method can effectively reduce the influence of image background noise, enhance the discrimination performance of visual words, and improve the performance of image classification based on Visual Language Model furthermore.