Research On Image Categorization Based On BOW And Visual Attention Model

In the age of "Big Data", people are no longer worried about how to get information. Instead, they feel anxious since data increases so fast. Nowadays, how to explore useful, meaningful information which users really care about is so valuable that many researchers pay attention to it. Among these numerous data, image data occupies a very important part. In fact, an image contains more information than many texts, and attracts people’s eyes easier. Moreover, people can understand image easier than text. Thus, it is meaningful to do research on large-scale image processing. Large-scale image processing includes several research fields such as image categorization, image retrieval and object detection, and all these fields cross each other. Image categorization, which is one of the basic problem in computer vision, can be widely used in different industry. Predictably, in the future, with the development of technology, large-scale image categorization will be used in daily life of people deeply and changes their life a lot. So it is really worthy to study large-scale image categorization.However, since a) the variation of light, scale and direction; b) different appearance in different period of the same kind of object; c) similar appearance of different object, it is hard to distinguish different images from various categories, so we usually need to represent these images into some special form to do classification. Recently, to solve it, many models have been proposed. Among these models, Bag-of-Words model(BOW) achieved very good result and played a really crucial role in the field. BOW model represents an image into a set of visual words which is quantization of local feature. Nonetheless, most algorithms based on BOW model look an image as an integrated object and count the frequency of features extracted from the whole image. Without distinguishing whether the features are extracted from the foreground or background of an image, those features from background brings a lot of disturbing for image categorization. By considering this disadvantage, this paper exploited visual attention model to solve it. Thus it is meaningful to do research on image categorization based on BOW model and visual attention model.In this paper, firstly we improved the algorithm of image categorization, and used visual attention model to distinguish the feature from foreground or background, then we proposed a new image categorization algorithm based on BOW model and visual attention model, which represents image more efficiently. After that, based on the proposed algorithm, we applied image categorization in ship detection of remote sensing and advertising recommendation system of image, and proposed a new ship detection algorithm which based on SEBOW model and a recommendation algorithm which based on BOW and visual attention model. The contributions are as follow:(1) Since conventional BOW model count the frequency of local feature in the whole image, each local feature plays the same role in the representation, the features from background of image influent the classification result. In order to solve it, this paper proposed a new algorithm based on BOW model and visual attention model. And it used visual attention model to describe foreground and background in the image, then mixed local feature and their saliency information to build codebook, thus the result of classification became better.(2) Since conventional ship detection sets a threshold for segmenting the ship from sea clutter, so it needs to pre-define the model of sea clutter, this method cannot apply to all situation, and it also ignore the shape and other information about ship. To solve this problem, this paper analyzed different distribution of features on the image, and distinguished them from whether foreground or background, then proposed a novel ship emphasized Bag-of-Words (SEBOW) model to improve the representation of SAR images consisting of ships and sea clutters. Based on the SEBOW model, this paper introduced a sliding window approach to detect ships on SAR image.(3) Conventional research on image ads is focus on how to use image labels and text surrounding the image to recommend ads, however text cannot describe image accurately, and it may ignore some important information in images. This paper used the BOW representation and visual attention information of images to analyze the similarity between images and the topics, then did the classification. After that, based on the idea of native ads, this paper proposed a new recommend algorithm based on image content to recommend ads. This recommend method won’t disturb user’s feeling and increase the click rate also.