Efficient Variable Bit-Rate Neural Image Compression with Perceptual-Enhanced Optimization For CLIC2025

Daxin Li

Abstract: In this paper, we propose a novel and enhanced image compression framework that builds upon the state-of-the-art (SOTA) DCVC-RT intra model, with a particular emphasis on advancing perceptual quality in compressed images. Although DCVC-RT demonstrates outstanding rate-distortion performance and real-time processing capabilities, it is still susceptible to generating perceptual artifacts, such as blurring and loss of fine textures, especially at lower bitrates. To effectively mitigate these issues, we introduce a comprehensive perceptual optimization strategy that leverages a semantic ensemble loss. This loss function is meticulously designed by integrating multiple complementary components, including Charbonnier loss for robust pixel-wise fidelity, perceptual loss to preserve high-level semantic features, style loss to maintain texture and style consistency, and a non-binary adversarial loss to further enhance the realism of reconstructed images. Our approach is developed as a solution for the CLIC2025 challenge, and we participate under the team name Vcoder. Through experiments, we demonstrate that our method significantly improves the perceptual quality of compressed images.

Team Name: Vcoder

Submission Number: 2