Ablah AlAmri¹ and Charith Abhayaratne², ¹The University of Sheffield, UK

Recently, learning-based image codecs have improved compression leading to excellent performance in bitrate reduction. However, their performance when transmitted over lossy channels has not been well studied. This paper investigates how these learning based image codecs perform under lossy channel transmission conditions. For this, we set up an experimental model that includes an encoder, a channel coding module, a channel simulation module and a decoder to evaluate the visual quality performance under various channel conditions. We compare the performance of several artificial intelligence models with the standard JPEG under various channel conditions and across various bitrates. The experimental results show that, under clean conditions, the learning-based codecs used in the experiments outperform JPEG in terms of PSNR and MS-SSIM. However, in a noisy channel, these codecs show significant degradation in PSNR and MS-SSIM under low-SNR conditions (especially below 12 dB SNR), whereas JPEG is more robust to channel errors and shows a more gradual degradation in quality as the SNR decreases.

channel errors, error robustness, JPEG AI, visual quality, learning-based image coding, AI-based image coding.