Fast implementation of SSIMULACRA2 for image quality assessment

Image quality assessment (IQA) plays a fundamental role in image processing and has undergone various developments. As IQA methods are specialized for image coding, SSIMULACRA2 exists, and these are utilized in recent image coding standards: JPEG XL, Jpegli, and AVIF. However, the new IQA of SSIMULACRA2 is computationally intensive, and even in reference implementations, the computation time can exceed that of the encoding process itself. In this paper, we propose a fast implementation of SSIMULACRA2. We achieve this by replacing convolution with separable FIR filters, skipping the computation of the low-contributed features, and separating the evaluation process for input and degraded images. The experimental results show that SSIMULACRA2 achieves a speedup of up to x82.4 compared to the original, with almost no degradation in correlation as measured by subjective evaluation.