FJNDF-Pytorch: Unified JND Pre-Filtering Benchmark

• FJNDF-Pytorch is an open-source unified benchmark and learning platform for frequency-domain JND pre-filtering, offering systematic evaluation and training for perceptual image coding.
• The platform integrates various mainstream JND models, supports multiple codecs and datasets, and provides objective BD-BR metrics for fair and reproducible comparisons.
• Its lightweight CNN pre-filter design and hybrid loss function significantly enhance computational efficiency while delivering superior rate-distortion performance.

FJNDF-Pytorch is an open-source unified benchmark and learning platform for frequency-domain Just Noticeable Distortion (JND)-guided pre-filtering in perceptual image coding. It provides a standardized environment to evaluate, compare, and train both traditional and neural pre-filtering methods, addressing critical challenges in computational efficiency and fair comparison through objective rates and reproducible pipelines (He et al., 12 Oct 2025).

1. Benchmark Architecture and Purpose

FJNDF-Pytorch was developed to resolve two foundational issues in frequency-domain JND pre-filtering: the reliance on subjective evaluation metrics and the significant computation overhead of prior models. The benchmark presents:

• A modular architecture integrating various mainstream JND models and injection strategies, supporting coefficient suppression and filter-based techniques.
• Native support for multiple open-source codecs—x264, x265, libaom, VVenC—and major public datasets including HEVC-B, XIPH, MCL-JCV, and MCL-JCI.
• Objective rate-distortion (BD-BR) metrics (e.g., PSNR, PSNR-HVSM, SSIM, MS-SSIM, VMAF, VMAF-NEG), enabling reproducible and quantitative method comparison.
• Two principal pipelines: (1) benchmarking for traditional JND injection algorithms and (2) data generation to supervise learning-based pre-filter development.

This structure enables rapid, consistent evaluation and data collection for subsequent neural network training.

2. Computational Efficiency and Rate-Distortion Performance

The platform facilitates substantial improvements in pre-filtering efficiency and compression results:

• Existing frequency-filter methods have high computational costs (e.g., up to 313% increased encoding time); FJNDF-Pytorch’s lightweight CNN model processes a 1080p image in 7.15 GFLOPs, merely 14.1% that of previous lightweight networks.
• Benchmarks reveal superior BD-BR savings across datasets and codecs. For example, on HEVC-B with VVenC encoding, the method achieves a VMAF BD-BR saving of –4.84%, outperforming other designs.
• Robust compression efficiency is maintained or improved across all tested codecs (libaom, x265, x264), with markedly lower resource requirements.

This positions FJNDF-Pytorch as a reference point for speed and coding gain in perceptual pre-filter design.

3. Lightweight CNN Pre-Filter Design and Training

The platform’s learning framework builds a purpose-designed, lightweight CNN for JND-guided pre-filtering:

• The network employs a supervised residual learning strategy, taking as target the residual between the unfiltered input and the ideal filtered reference produced by top-performing JND methods. This “residual distillation” simplifies the learning target and enables the CNN to outperform its reference.
• Architectural modifications include replacing 5×5 with 3×3 convolutions to better capture DCT-domain locality and a bottleneck-expansion channel topology (Head, Body, Tail stages) for parameter efficiency.
• The inference model contains only 1.86K parameters.

The learning pipeline utilizes over 2,000 original-reference pairs from datasets like DIV2K and KonJND-1K. High-quality supervision is ensured via the standardized training pipeline provided by the benchmark.

4. Hybrid Loss Construction and Frequency-Domain Enforcement

A central innovation is the hybrid loss function, ensuring simultaneous spatial and frequency-domain fidelity:

• The full objective is:

$$L_{all} = \lambda_1 L_c(I_f, I_{gt}) + \lambda_2 L_m(I_f, I_{gt}) + \lambda_3 L_{freq}(C_f, C_{gt}, C_o)$$

with $L_c$ a Charbonnier pixel loss, $L_m$ the MS-SSIM structural similarity, and $L_{freq}$ consisting of: - $L_{dct}^{res}$: residual-based distillation of DCT coefficients - $L_{dct}^{cons}$: conservation constraint partitioning the $8 \times 8$ DCT block into low/high frequency regions; this term enforces preservation of low-frequency content while suppressing (but never amplifying) high frequencies.

The quantitative JND model, used for injection and suppression, is:

$$J_T(u, v, N) = s \cdot J_{CSF} \cdot J_{LA} \cdot J_{CM} \cdot J_{SA}$$

where $(u,v)$ label DCT frequencies within $N \times N$ blocks, and $s$ aggregates summation effects across the underlying perceptual dimensions.

Actual JND-guided suppression is performed via:

$$C_f = \begin{cases} \text{sgn}(C_o) \cdot \sqrt{ C_o^2 - p(u,v,B) \cdot J_T^2}, & \text{if } |C_o| \geq J_T \ 0, & \text{otherwise} \end{cases}$$

where $p(u,v,B)$ is a suppression weight.

5. Experimental Validation and Comparative Analysis

All protocols employ standardized datasets (HEVC-B, XIPH, MCL-JCV, MCL-JCI) and public codecs for benchmarking and evaluation. The method demonstrates:

• Superior BD-BR rate-distortion gains (e.g., VMAF BD-BR saving of –4.84% on HEVC-B, with further datasets showing consistent quality and robustness).
• Dramatically reduced computational cost compared to prior lightweight (as well as conventional) pre-filtering approaches.
• Competitive or improved perceptual quality versus the best published alternatives (including those by Kang et al., Tan et al., and Sun et al.), without sacrificing objective fidelity.

Results are reported using exclusively objective metrics, circumventing labor-intensive subjective scoring.

6. Reproducibility and Standardization in Research

FJNDF-Pytorch’s open-source release (https://github.com/viplab-fudan/FJNDF-Pytorch) serves to unify future research in JND-guided perceptual image coding:

• Direct, standardized comparison across methods and datasets fosters reproducibility.
• The benchmark environment covers the full research workflow: data preprocessing, model zoo, injection strategy selection, codec interfacing, and downstream neural training and evaluation.
• Objective, rate-distortion-centric evaluation further ensures that compression efficiency estimates are reproducible and comparable across subsequent studies.

This standardized platform suggests a plausible direction towards fair, automated benchmarking in the field, lowering methodological barriers for both neural and non-neural model development.

7. Context and Significance

The introduction of FJNDF-Pytorch addresses pressing obstacles in perceptual image coding—namely, unfair or inconsistent benchmarking and excessive computational cost. Its deployment of a computationally parsimonious, frequency-domain-aware neural architecture sets a new reference point for practical, high-performance perceptual coding tools. The reproducible design and open-source accessibility imply that future research in JND-driven pre-filtering will likely converge around objective metrics and lightweight implementations enabled by this platform.