Optimizing Multistage Discriminative Dictionaries for Blind Image Quality Assessment

Abstract

According to the types of the used features, state-ofthe-art methods for blind image quality assessment (BIQA) typically have two categories. The first category utilizes handcrafted natural scene statistics (NSS) features which are derived based on the statistical regularity of natural images. The second category utilizes codebook-based features which are obtained by feature encoding over a pre-constructed codebook. However, several problems need to be addressed in the existing codebook-based BIQA methods. First, dimension of the involved codebook-based features is too high. Second, codebooks are usually constructed by unsupervised learning algorithms. To address these problems, we propose a novel codebook-based BIQA method by optimizing multi-stage discriminative dictionaries (MSDDs). To be more specific, MSDDs are learned by performing the label consistent K-SVD (LC-KSVD) algorithm in a stage-by-stage manner. In each stage, a new quality consistency constraint called “qualitydiscriminative regularization” term is introduced and incorporated into the traditional reconstruction error term to form a unified objective function which can be effectively solved by LC-KSVD. Then, the latter stage takes the reconstruction residual data in the former stage as input based on which LC-KSVD is repeatedly performed until the final stage is reached. Once the MSDDs are learned, multi-stage feature encoding (MSFE) is performed to extract feature codes. Finally, the feature codes are concatenated across all stages and spatially aggregated over the entire image for quality prediction via a typical regression module. The proposed method has been extensively evaluated on five IQA databases and the experimental results has well confirmed its superiority over the existing relevant BIQA methods.