Abstract:This paper describes a novel optimization framework for visual codebook generation. Firstly, the Condition Number (CN) is applied to evaluate the stability of initial visual features, and the well conditioned features are preserved by eliminating the bad ones. At the mean time, an adaptive algorithm to generate low-dimensional visual words is proposed by studying the relationship between clustering and dimension-reducing. In order to overcome the popular LBG codebook design algorithm suffers from local optimality and is sensitive to the initial solution, a parameter called neighborhood-support for each feature is calculated according to clustering structure, which is used to select initial visual words adaptively. Finally, the rational distortion function is redefined using Silhouette. Compared with traditional algorithm, the presented algorithm has excellent properties at simultaneous clustering and dimension reduction, good robustness and adaptive optimization. A good performance (73.46% classification rate) of application this method to 13-Scene classification is obtained by using Probabilistic Latent Semantic Analysis (PLSA).
杨丹, 李博, 赵红. 鲁棒视觉词汇本的自适应构造与自然场景分类应用[J]. 电子与信息学报, 2010, 32(9): 2139-2144.
Yang Dan, Li Bo, Zhao Hong. An Adaptive Algorithm for Robust Visual Codebook Generation and Its Natural Scene Categorization Application. , 2010, 32(9): 2139-2144.