Image Quality Assessment: Exploring the Similarity of Deep Features via Covariance-Constrained Spectra
Shujun Lang, Mingliang Zhou, Xuekai Wei, Jielu Yan, Yong Feng, Weijia Jia
Abstract
In this paper, we propose a full-reference image quality assessment (FR-IQA) method based on covariance-constrained spectral derivation, which achieves degradation perception of distorted images through a designed perceptual deep feature similarity strategy. First, original-resolution images were adopted as inputs to preserve information invariance, whereas fundamental feature extractors were employed to obtain both shallow and deep-level image features. Second, guided by statistical principles, centralized preprocessing was conducted to enhance data stability, covariance constraints were utilized for feature relationship characterization, joint spectral analysis was implemented for feature–pair information unification, and eigenvalue analysis was performed to complete spectral derivation. Finally, the spectral derivation results from both feature levels were integrated to compute the predicted quality scores for distorted images. The experimental results demonstrated that our method exhibited comparable performance and competitiveness with state-of-the-art approaches.