Performance Evaluation of Learning based Image Coding Solutions and Quality Metrics
The JPEG AI ad hoc group aims to evaluate the performance of deep learning-based image coding solutions with a precise and well-defined evaluation methodology. Thus, subjective image quality assessment experiments were conducted during the 84th JPEG meeting in Brussels, Belgium, by a mix of experts and non-expert observers (18 total). The experiments evaluated the performance of five learning-based image coding solutions against four transform-based codecs (HEVC, WebP, JPEG2000 and JPEG), on 8 SD to UHD natural images, at four different bitrates. The experimental results obtained and reported during the 85th JPEG meeting in San Jose, California show that subjective and objective qualities of state-of-the-art learning-based image coding algorithms were competitive to transform-based codecs. Thorough inspection on the visual results revealed the typical artifacts encountered in the learning-based codecs. Moreover, several full-reference objective quality metrics were evaluated to find which metric correlates better with human opinion scores, for different types of coding solutions, i.e. for transform and learning-based image codecs.
The full “Performance Evaluation of Learning based Image Coding Solutions and Quality Metrics” report is available for download here.