Multidimension Attention Network for Full-Reference Light Field Image Quality Assessment

Light Field (LF) imaging enables the post-capture adjustments in focus and view, which facilitates many new applications. However, various types of distortions and noises are inevitably introduced during LF image generation and processing procedures. It is crucial to assess the perceptual quality of distorted LFIs to promote the development of LF imaging. Most existing full-reference LF quality assessment methods are inherently limited as they overlook the global and local dependencies of the cropped image patches. To mitigate this issue, this paper proposes a multi-dimension attention network for full-reference LF image quality assessment. Specifically, we first propose a multi-level Chebyshev stack structure that reorganizes LF views into auxiliary view stacks under the same Chebyshev distances along horizontal, vertical, left- and right-diagonal directions. By constraining the Chebyshev distances, the view correlation can be kept consistent in each angular direction. Then, a multi-dimension attention network is put forward, which consists of three components including the vision Transformer (ViT) feature extraction module, the multi-dimension discrepancy transformer module, and the quality regression module. By aggregating the ViT and Swin Transformer, the proposed network is able to allow an adequate information interactions of the reference, distorted, and the calculated discrepancy features globally and locally. Comprehensive experiments on three public LF image quality evaluation datasets illustrate the superiority of the proposed method. The code is available at https://github.com/ldyorchid/LF-IQA.