A Feature Encoding Based on Fuzzy Codebook for Large-Scale Image Recognition

Feature encoding is an important step in image representation pipeline. In recent works, a codebook approach is generally used for image representation, and image is represented by codebook based encoding of extracted local descriptors. The Fisher Vector (FV) encoding is known as a powerful technique, however the memory cost of the codebook is rapidly increased dependence on the dimensionality of local feature. This paper presents a feature encoding technique based on compact codebook by using fuzzy clustering. Our approach contains two ideas. The first idea is to modify the fuzzy clustering for suitable calculation of high-dimensional vectors such as a local feature. The second idea is to update the codebook to each image, and embed its difference to image feature by using KL divergence. Our approach recorded 1.04% higher rate than FV on Caltech-101 dataset. The advantage of our approach is that the model parameter of the codebook is small.