Contrastive Learning Feature Enhancement and High–Low Frequency Texture Interaction Networks for DIBR-Synthesized View Quality Assessment

Depth image-based rendering (DIBR) is a common method for synthesizing virtual views to achieve smooth transitions in immersive media, but its immature technology often introduces distortions, adversely affecting visual quality. Obviously, accurately assessing the quality of synthesized views is crucial for monitoring and guiding rendering process. To this end, this paper proposes a no-reference deep learning-based quality assessment method for DIBR-synthesized views, which is primarily achieved by combining a contrastive learning feature enhancement network and a high-low frequency texture interaction network, abbreviated as CONTIN. Different from the traditional methods based on handcrafted feature extraction, the proposed method employs an end-to-end deep learning approach, fully exploiting the data characteristics and feature correlations. Specifically, to address the issue of sample expansion in existing deep learning methods, a contrastive sample database is first constructed by simulating various traditional and rendering distortions based on natural images, and training is performed on this database to obtain a contrastive learning feature enhancement network, which is used to extract contrastive features. Additionally, since contrastive learning tends to focus on learning abstract semantic-level features rather than pixel-level texture details, a wavelet transform decoupling is further applied to the synthetic distortion samples to construct a high-low frequency texture interaction network for extracting texture features. Finally, the two types of features are fused and regressed to generate the final quality score. Experimental results show that the proposed method achieves superior performance across three benchmark databases (namely IRCCyN/IVC, IETR and MCL-3D), with PLCC reaching 0.9404, 0.8380, and 0.9666 respectively, representing improvements of 0.0179, 0.0350, and 0.0175 higher than the existing best methods.