Supervised dictionary learning for blind image quality assessment using quality-constraint sparse coding

• We propose a new general purpose OF-BIQA model without training on the human subjective scores.
• We propose a supervised dictionary learning framework by using quality-constraint sparse coding.
• Besides Gabor features, we also incorporate HOG descriptor for local feature representation.

Blind image quality assessment (BIQA) involves predicting the perceptual quality of distorted images without using their corresponding reference images as benchmark. Especially, it is desirable and meaningful to design effective opinion-free BIQA (OF-BIQA) model to predict image quality without depending on human subjective score. Toward this end, we propose a supervised dictionary learning framework for OF-BIQA using quality-constraint sparse coding. The prominent advantage of the proposed model is that “ground truth” quality scores derived from existing full-reference IQA (FR-IQA) metrics are incorporated into the traditional dictionary learning framework so that a quality-aware sparse model can be learnt. Since the goal of BIQA is to predict the quality score, the introduction of quality information into dictionary learning can be regard as a supervised dictionary learning framework. In the detailed implementation, a quality-aware regularization term is added to the traditional dictionary learning formulation, such that a feature-aware dictionary and a quality-aware dictionary can be learned jointly. Especially, these two dictionaries share the same sparse coefficients, so that the reconstruction errors from the image feature vectors and quality score vectors are both minimized. Once the feature-aware and quality-aware dictionaries are jointly learned, given a testing sample, we first abstract its feature vector and then compute the corresponding sparse coefficients w.r.t. the learnt feature-aware dictionary, finally its quality score can be directly reconstructed based on the learnt quality-aware dictionary and the estimated sparse coefficients w.r.t. the learnt feature-aware dictionary. The reconstructed quality score is expected to well approximate to the “ground truth” quality score. Thorough validation experiments on three publicly available IQA benchmark databases demonstrate the promising performance of the proposed OF-BIQA model both on the prediction accuracy and generalization capability.