关键词: 图像分类, 视觉词袋模型, 视觉词典, 谱聚类

Abstract: Construction of visual vocabulary is a crucial step in popular Bag-of-Visual-Words（BOVW） model. Currently, K-means clustering is generally applied to constructing the visual vocabulary. However, the visual dictionary tends to be of low discrimination due to limitation of K-means clustering and complexity of high dimensional spatial structure of samples. Under the frame of spectral clustering, a dictionary learning algorithm with stronger discriminative capability is proposed. In order to reduce degradation of descriptors discriminative power during quantization and the inherent problems of storage and calculation in spectral clustering, the training samples are divided into sub-sample sets according to the label information of category. Centers of each data set are obtained based on spectral clustering with Nystr?m algorithm and then the final compact visual dictionary is generated. Experimental results in Scene-15 dataset verify the correctness and effectiveness of the proposed algorithm. Especially when the training samples are limited, the visual dictionary via the algorithm can obtain better performance.

Key words: image classification, bag of visual words, visual vocabulary, spectral clustering