Category-dependent learned image compression for smartphone photography with standard-compliant decoders

mso-fareast-font-family:Calibri;mso-fareast-theme-font:minor-latin;mso-ansi-language:
FR;mso-fareast-language:EN-US;mso-bidi-language:AR-SA">In this paper, we address the need for tailored image compression by focusing on the specific use case of smartphone photography, particularly selfie, food and landscape images, which dominate user-captured photos. We adapt SegPIC by fine-tuning it on a dedicated selfie, food and landscape dataset while keeping the decoder unchanged to maintain compatibility with JPEG AI standard decoding requirements. The model's performance was evaluated using Kodak dataset and the JPEG AI test set, with comparisons based on PSNR and MS-SSIM metrics. To assess the generalization of category-specific fine-tuning, we evaluated the MBT model across the same categories. Our results demonstrate that this fine-tuning improves compression efficiency and image quality compared to training on general datasets, highlighting the benefits of category-specific training within standardized frameworks.