Automatic Acquisition of Appropriate Codewords Number in BoVW Model and the Corresponding Scene Classification Performance

Abstract

Scene classification has received significant attention in past few years, particularly in the scene semantic comprehension. The Bag of Visual Words (BoVW) is one of the most widely used models in current scene classification research. Acquisition of appropriate codewords number in BoVw Model is a key step in BoVw model. The classical framework to acquire the appropriate number of codewords needs to cluster the codewords by K-Means for many iterations, rendering its high computation cost. In the current study, an improved framework was proposed to acquire the appropriate number of codewords automatically, and corresponding scene classification performance with Latent Dirichlet Allocation (LDA) topic model was investigated. The experimental results demonstrate the core capability of proposed framework. The computation cost of scene semantic comprehension can be reduced greatly by means of the proposed framework, particularly in amount sequence image processing such as CT, MRI and other medical image analysis.