A Quantitative Quality Measurement for Codebook in Feature Encoding Strategies

抄録

Nowadays, a feature encoding strategy is a general approach to represent a document, an image or audio as a feature vector. In image recognition problems, this approach treats an image as a set of partial feature descriptors. The set is then converted to a feature vector based on basis vectors called codebook. This paper focuses on a prior probability, which is one of codebook parameters and analyzes dependency for the feature encoding. In this paper, we conducted the following two experiments, analysis of prior probabilities in state-of-the-art encodings and control of prior probabilities. The first experiment investigates the distribution of prior probabilities and compares recognition performances of recent techniques. The results suggest that recognition performance probably depends on the distribution of prior probabilities. The second experiment tries further statistical analysis by controlling the distribution of prior probabilities. The results show a strong negative linear relationship between a standard deviation of prior probabilities and recognition accuracy. From these experiments, the quality of codebook used for feature encoding can be quantitatively measured, and recognition performances can be improved by optimizing codebook. Besides, the codebook is created at an offline step. Therefore, optimizing codebook does not require any additional computational cost for practical applications.