Chrominance Aware Progressive Rate-Distortion Optimization for Learned Image Compressio

We present an end-to-end learned image compression system trained under a unified rate-distortion objective, comprising a convolutional encoder, a U-Net-based entropy model, and a convolutional decoder operating in the YCbCr colourspace. The entropy model predicts per-symbol Gaussian distribution parameters over the quantized latent representation, providing a differentiable bitrate surrogate during training, while inferencetime compression is performed using zlib on the serialized latent. We introduce a progressive rate-distortion training strategy in which multiple model variants are obtained through iterative finetuning across successive stages, each adapting the rate-distortion balance without independent retraining. Additionally, we propose a chrominance-aware distortion curriculum that progressively reallocates channel weighting from luminance-dominant to balanced YCbCr emphasis, enabling stable structural reconstruction prior to enhanced colour fidelity. The proposed system achieves compression ratios of approximately 32×, 17×, and 3.2× relative to uncompressed PNG, with PSNR values of up to 32.78 dB on BSD100, while maintaining total encode-to-decode latency under 35 ms. Evaluation across BSD100, Set14, and Urban100 demonstrates consistent performance across diverse image content.