Introduction
Traditional image compression relies on discrete quantization and entropy coding, processes that are inherently non-differentiable. This poses a significant challenge for end-to-end training of deep learning models, primarily due to the "zero-gradient problem" where the derivative of the quantization staircase function is zero almost everywhere. Furthermore, accurately estimating the bit-rate cost of complex entropy coders like Context-Adaptive Binary Arithmetic Coding (CABAC) during training is difficult. This paper introduces a "soft bit" representation to bridge this gap, enabling differentiable rate-distortion optimization.
System Architecture
The proposed framework integrates a non-linear autoencoder with a context-adaptive bit-plane image coding (CABIC) module. The architecture includes an encoder that extracts features, which are then processed through a soft quantization and bit-plane decomposition stage. This allows the system to estimate the actual bit-rate cost during the training phase.
Methodology
Soft Bit Conversion
To overcome the non-differentiability of quantization, the authors represent quantized values as a sequence of "soft bits." Instead of hard binary values, each bit is approximated by a superposition of sigmoid functions. As shown in the figure below, this transformation makes the bit representation a differentiable function of the input feature, allowing gradients to flow back to the encoder during backpropagation.
Context-Adaptive Bit-plane Image Coding (CABIC)
Inspired by JPEG2000, the system organizes bits into bit-planes. A context model predicts the probability of the current bit based on its spatial neighbors. To make this process differentiable during training, a "Probability Regressor" network is employed to estimate the probability of the soft bits. This allows the model to optimize for the actual bit-rate achieved by the arithmetic coder, rather than relying on crude proxies like added noise.
Experimental Results
The model was trained on the CLIC 2018 dataset and evaluated on the Kodak PhotoCD benchmark.
- Performance: The method achieves state-of-the-art MS-SSIM among learning-based schemes, performing comparably to BPG (HEVC-based) and significantly outperforming JPEG and JPEG2000.
- Bit Allocation: Visual analysis confirms the model learns to prioritize complex regions, effectively "zeroing out" less important bit-planes at lower bit-rates to optimize storage while maintaining structural integrity.
Conclusion
By introducing differentiable soft bits and a learned rate estimator, this work successfully bridges the gap between discrete compression and continuous optimization. While a small PSNR gap remains compared to top-tier traditional codecs like BPG, the approach offers a robust path for high-perceptual-quality learned image compression with realistic rate estimation.
Summary updated