9 October 2023 in IEEE Transactions on Circuits and Systems for Video Technology
Quantizing a floating-point neural network to its fixed-point representation is crucial for Learned Image Compression (LIC) because it improves decoding consistency for interoperability and reduces space-time complexity for implementation. Existing solutions often have to retrain the network for model quantization, which is time-consuming and impractical to some extent. This work suggests using Post-Training Quantization (PTQ) to process pretrained, off-the-shelf LIC models. We theoretically prove that minimizing quantization-induced mean square error (MSE) of model parameters (e.g., weight, bias, and activation) in PTQ is sub-optimal for compression tasks and thus develop a novel Rate-Distortion (R-D) Optimized PTQ (RDO-PTQ) to best retain the compression performance. Given a LIC model, RDO-PTQ layer-wisely determines the quantization parameters to transform the original floating-point parameters in 32-bit precision (FP32) to fixed-point ones at 8-bit precision (INT8), for which a tiny calibration image set is compressed in optimization to minimize R-D loss. Experiments reveal the outstanding efficiency of the proposed method on different LICs, showing the closest coding performance to their floating-point counterparts. Our method is a lightweight and plug-and-play approach without retraining model parameters but just adjusting quantization parameters, which is attractive to practitioners. Such an RDO-PTQ is a task-oriented PTQ scheme, which is then extended to quantize popular super-resolution and image classification models with negligible performance loss, further evidencing the generalization of our methodology. Related materials will be released at https://njuvision.github.io/RDO-PTQ.
14 November 2025 in IEEE Journal of Selected Topics in Signal Processing
Hyperspectral image (HSI) signals are high-dimensional, continuous, and sensor-variant, posing significant challenges for compression and restoration-especially under the scarcity of labeled data and the limitations of discretized, supervised models. In this work, we propose HINER++, a unified, self-supervised framework based on implicit neural representations (INRs) for continuous hyperspectral signal modeling. HINER++ formulates each HSI sample as a wavelength-to-intensity mapping, parameterized by a compact neural network. Thus, HSI signal modeling becomes equivalent to a continuous neural function regression problem, where compression is achieved by encoding the parameters of the neural function. Building on the principles of Deep Image Prior (DIP), we further reinterpret the inductive bias of the INR architecture as a natural prior for HSI restoration, enabling tasks such as denoising, inpainting, and super-resolution without labeled supervision. Finally, we investigate the impact of lossy INR-based compression on downstream perception tasks, using classification as a proxy, and demonstrate that such implicit modeling helps preserve task performance under extreme compression via proposed task-aware components-Adaptive Spectral Weighting (ASW) and Implicit Spectral Interpolation (ISI). Extensive experiments on real-world HSI benchmarks demonstrate that HINER++ achieves superior performance across multiple compression and restoration tasks, preserving classification accuracy under extreme compression ratios (up to 100×). This work offers a new perspective: compression is not merely data reduction, but a self-supervised process of signal disentanglement and recovery. All source code and data will be made publicly available at https://github.com/Eric-qi/HINER to support reproducible research.
1 November 2026 in IEEE Transactions on Circuits and Systems for Video Technology
1 January 2026 in IEEE Journal on Emerging and Selected Topics in Circuits and Systems