Scoring structure regularized gradient boosting network for blind image quality assessment

Abstract

Blind image quality assessment (BIQA) aims to quantitatively predict the subjective perception of the distorted image without accessing its corresponding clean version. Prevailing methods typically model BIQA as a regression task and strive to minimize the average prediction error in terms of the pointwise unstructured loss, such as Mean Square Error (MSE) or Mean Absolute Error (MAE), which ignores the perception toward the rank orders and perceptual differences between different images. This paper proposes a Scoring Structure regularized Gradient Boosting Network (SSGB-Net) to achieve a more comprehensive perception across all distorted images. More specifically, our SSGB-Net performs BIQA in three stages, pair-wise rectification and list-wise boosting, followed by point-wise prediction after linear transformation. First, we correct the initial scores by incorporating the structured pairwise loss, i.e., SoftRank, to preserve the perceptual rank orders of pairwise images. Then, we further boost the previous pairwise correction results with structured listwise loss, i.e., Norm-in-Norm, to maintain the perceptual difference across all images. Finally, the point-wise prediction measures the MSE between the transformed scores and the ground truth through a closed-form solution of the Exponential Moving Average (EMA) driven linear transformation. Based on these iterative corrections, our SSGB-Net can effectively balance multiple BIQA objectives and outperform many state-of-the-art methods in terms of Pearson Linear Correlation Coefficient (PLCC), Spearman Rank Correlation Coefficient (SRCC) and Root Mean Squared Error (RMSE).