Just Noticeable Distortion (JND)-guided pre-filter is a promising technique for improving the perceptual compression efficiency of image coding. However, existing methods are often computationally expensive, and the field lacks standardized benchmarks for fair comparison. To address these challenges, this paper introduces a twofold contribution. First, we develop and open-source FJNDF-Pytorch, a unified benchmark for frequency-domain JND-Guided pre-filters. Second, leveraging this platform, we propose a complete learning framework for a novel, lightweight Convolutional Neural Network (CNN). Experimental results demonstrate that our proposed method achieves state-of-the-art compression efficiency, consistently outperforming competitors across multiple datasets and encoders. In terms of computational cost, our model is exceptionally lightweight, requiring only 7.15 GFLOPs to process a 1080p image, which is merely 14.1% of the cost of recent lightweight network. Our work presents a robust, state-of-the-art solution that excels in both performance and efficiency, supported by a reproducible research platform. The open-source implementation is available at https://github.com/viplab-fudan/FJNDF-Pytorch.
Figures & tables
Fig. 1: Illustration of our main contributions. (a) We build a standardized benchmark for frequency-domain JND pre-filters. (b) Using this benchmark, we generate a dataset to train our proposed lightweight CNN via a supervised pipeline, which is optimized by a joint spatial and frequency loss function.
Fig. 2: An overview of the FJNDF-Pytorch framework. (a) The modular toolbox, whose components are detailed in Table I , supports two primary pipelines. (b) A benchmarking pipeline for evaluating filters and generating training data. (c) A training pipeline that leverages this data to train learning-based pre-filters.
TABLE I: Components Integrated in the FJNDF-Pytorch Toolbox
Dataset
Method
PSNR
PSNR-
MS-
VMAF-
ALL
M a + I b
HVSM
SSIM
NEG
HEVC-B [ 14 ]
[ 5 ] + [ 6 ]
1.94
-2.83
-1.36
-0.86
-0.78
[ 13 ] + [ 6 ]
1.95
-2.75
-1.35
-0.67
-0.70
[ 5 ] + [ 7 ]
4.91
-4.29
-2.04
-0.85
-0.57
XIPH [ 15 ]
[ 12 ] + [ 7 ]
-0.27
-4.17
-2.76
-2.87
-2.52
[ 13 ] + [ 6 ]
-0.45
-3.95
-3.14
-2.53
-2.52
TABLE II: Results of the top-three performing algorithms on VVenC
Fig. 3: Training pipeline of our lightweight pre-filter network. Our core contribution, the hybrid loss, supervises the network. The spatial-domain loss ( Lc,Lm ) is computed against a Ground Truth (GT) from a reference filter. The frequency-domain loss imposes a residual constraint ( Ldctres ) against the GT and a conservation constraint ( Ldctcons ) against the original input. The network backbone (MBR, FST and HDPA) is detailed in [ 33 ] . MBR is re-parameterized into standard convolutions at inference pipeline.
Fig. 4: Illustration of the frequency area partitioning for our proposed conservation constraint loss, Ldctcons . An 8x8 DCT block is divided into a Low-Frequency ( LF ) and a High-Frequency ( HF ) area based on a zigzag scan order and a predefined threshold, K .
Dataset
PSNR
PSNR-HVSM
SSIM
MS-SSIM
VMAF
VMAF-NEG
ALL
[ 5 ]
[ 6 ]
[ 9 ]
Ours
[ 5 ]
[ 6 ]
[ 9 ]
Ours
[ 5 ]
[ 6 ]
[ 9 ]
Ours
[ 5 ]
[ 6 ]
[ 9 ]
Ours
[ 5 ]
[ 6 ]
[ 9 ]
Ours
[ 5 ]
[ 6 ]
[ 9 ]
Ours
[ 5 ]
[ 6 ]
[ 9 ]
Ours
HEVC-B
1.55
1.95
1.51
0.87
-1.81
-2.75
-2.53
-4.20
1.84
-0.73
-1.78
-2.05
-0.79
-1.35
-2.34
-2.90
1.50
-0.54
-2.42
-4.84
0.03
-0.67
-1.81
-3.67
0.39
-0.68
-1.56
-2.80
XIPH
-0.77
-0.45
-0.31
-1.16
-3.15
-3.95
-3.07
-5.32
0.82
-2.15
-1.98
-2.78
-1.66
-3.14
-2.78
-4.13
0.09
-2.55
-4.12
-6.26
-1.31
-2.53
-3.27
-5.07
-1.00
-2.46
-2.59
-4.12
MCL-JCV
0.45
0.32
0.21
-0.26
-2.00
-2.57
-1.94
-3.64
1.02
-1.21
-1.30
-1.73
-1.33
-2.07
-1.88
-2.69
1.58
-1.38
-2.72
-5.05
0.21
-1.50
-2.12
-3.61
-0.01
-1.40
-1.63
-2.83
MCL-JCI
2.81
4.81
3.73
2.74
-6.49
-6.31
-7.51
-8.80
3.01
-0.17
-2.04
-2.20
-3.20
-2.77
-5.57
-5.98
-0.91
-2.58
-5.62
-7.55
-2.66
-2.77
-4.88
-6.55
-1.24
-1.63
-3.65
-4.72
TABLE III: Performance comparison of different pre-filter methods across various datasets on VVenC.
Encoder
Method
PSNR
PSNR-
MS-
VMAF-
ALL
HVSM
SSIM
NEG
libaom
Ours
-0.34
-5.70
-4.71
-6.20
-4.24
[ 9 ]
0.67
-3.26
-3.49
-3.29
-2.34
[ 6 ]
1.42
-3.25
-2.14
-1.48
-1.36
[ 5 ]
1.30
-2.43
-1.26
0.66
-0.43
x265
Ours
0.00
-6.81
-4.47
-5.66
-4.23
TABLE IV: Performance comparison of different pre-filter methods across various encoders on HEVC-B.
Method
PSNR
PSNR-HVSM
MS-SSIM
VMAF-NEG
ALL
Reference
1.95
−2.75
−1.35
−0.67
−0.70
BaseLine
3.52
−5.50
−3.50
−2.85
−2.08
BaseLine + Lfreq
0.69
−4.43
−3.34
−4.18
−2.81
Ours (Lightweight)
0.87
−4.20
−2.90
−3.67
−2.47
TABLE V: Ablation study of different components of HEVC-B on VVenC.
Method
Parameters (K)
GFLOPs
Inference Time
Train
Inference
CPU a (s)
GPU b (ms)
[ 5 ]
/
/
/
∼70
/
[ 6 ]
/
/
/
∼20
/
[ 9 ]
12.23
12.22
50.70
0.835
15.43
BaseLine
38.92
3.06
12.15
0.595
13.04
Ours
24.95
1.86
7.15
0.428
10.88
TABLE VI: Model size, computational cost, and inference latency.