Research Of Image Classification Based On Improved Bag-of-Features Model

As we all know, there are a vast number of images coming out every day due to the development of digital image technology and computer technology. Therefore automatic image classification is becoming more and more important because it is a basis to solve many other image problems. When we are faced with image processing problems, local features such as sift features usually provide good information for us, however the number of local features extracted from an image is usually very large and different form others, so it will cost a lot to use local features for image classification by traditional methods. Fortunately, Bag-of-Features model can use local features to represent image efficiently and it have already been used widely in the field of image retrieval, image classification, image recognition, and so on. However, Bag-of-Features model also have some deficiency, this paper will improve it in the following two aspects:Firstly, to construct feature vectors, one local feature is mapped to the most similar visual word in the traditional Bag-of-Features model. However, a local feature may be highly similar to many visual words at the same time. So, in order to increase the stability of Bag-of-Features model and make the performance of the model much better, this paper maps one local feature to several visual words to improve the accuracy of image classification.Secondly, different visual words may have different occurrence in one image, that is to say some visual words may appear randomly in all images, while others appear more often in some categories than in the others, so visual words may have different importance for image classification. So, feature vectors should be weighted, and the more important the words are, the higher weight they should gain. The most popular feature weighting algorithm in Bag-of-Features model is TF-IDF, however TF-IDF just computes the frequency of visual words and images and neglects the relationship between visual words and image categories, as a result it has some deficiency. This paper adopts chi square test algorithm which is computed based on the image categories to offset the deficiency of TF-IDF algorithm and make it performing better, so that the performance of image classification will be better by Bag-of-Features model.