Research On Single-scale Image Classification Employing Bag-of-Words Model

With the development of computer and Internet technology, vast amount of images spring up, which turn the ways of efficient management and utilization into an urgent task for researchers. As one kind of digital image processing technology, the automatic classification of images becames the break-through to address the problems mentioned above.However, the automatic classification of images bothers the academia about the development of image annotation and image retrieval all the time. The manual classification has still been applied in a lot of image classification systems. As the result of booming images on Internet, the manual classification brings several problems, such as large amount of tedious work, eye fatigue and inefficiency of staff, and classification error caused by ambiguous semantics. Therefore, the automatic classification of images gets more and more attention. Computer vision theory, which will be replacing the artificial image classification, not only improves the efficiency and reduces the labor intensity of personnels, but also makes the classification results more objective.This paper, studying the automatic classification of images, focuses on the analysis of the bag-of-words model, which is widely used in current image classification. In terms of the traditional ways of representing the features of the bag-of-words model, to construct multi-scale space while collecting SIFT features, usually causes huge computational complexity and insufficient description for a part of feature points. For the above shortcomings mentioned, this paper proposes a method of extracting SIFT feature points directly through the grid and describing it using the bag-of-words model for image classification, instead of building multi-scale space. This method of using the grid to fix the feature points’ extractive position, contributes to reducing the amount of computation and classifying SVM for taking into account the global of local features which give more obvious class differences among the visual words after forming the visual vocabulary. Experiments prove that bag-of-words model with single-scale SIFT description dose better in aspects of recall ratio, precision and aggregative indicator than regular one.Due to the low classifying efficiency of single-scale SIFT, what matters is the relatively high dimension of descriptors, which affects the retrieval efficiency to some extent. Thus, a method based on low dimension extracting from the single-scale image is bring forward for bag-of-words model. Regarding the final effect, it works the same as the single-scale bag-of-words model in aspects of recall ratio, precision ratio and aggregative indicator, while much more computational complexity is avoided during the process of creating the descriptor.