JPEG Processing Neural Operator for Backward-Compatible Coding

This repository contains the offical source code of our ICCV 2025 paper:

JPEG Processing Neural Operator for Backward-Compatible Coding

---

1. Overview

JPNeO inserts two lightweight neural operators around the standard JPEG codec:

Module	Role	Status
JENO	Encoder-side operator that learns a perceptual re-mapping before quantization	pre-trained weights provided
JDNO	Decoder-side operator that removes JPEG artifacts without IDCT	train from scratch
Q θ	Learned quantization matrices shared by encoder / decoder	pre-trained weights provided

The resulting .jpg bit-streams remain bit-exactly compatible with every legacy JPEG decoder.

---

1. Installation

Our code is based on Ubuntu 20.04, pytorch, CUDA (NVIDIA RTX 3090 24GB, sm86) and python.

For enviorment setup, we recommend to use CONDA for the installation:

conda env create --file env.yaml
conda activate jpneo

This DEMO file includes the pre-trained checkpoint of JPNeO’s encoder, JENO, along with the pre-trained quantization matrix ./bin/learned_qmaps.pt, enabling reviewers to directly test image compression. To test proposed model, modify each file's image path such as :

---

2. Training Decoder

Our JDNO is too heavy to be uploaded. Therefore, we provide the training code below.

The basic train code are following:

Training JDNO python train.py --config configs/train_JDNO.yaml --gpu 0

This process automatically generates a save folder and checkpoint files.

The pre-trained weights will be uploaded sooner or later.

---

4. Quick Demo (Image-level Compression)

You can test the pre-trained JENO and learned quantization matrix using the provided script `TRYME.ipynb'.

Then run the script to compress the image using JENO and $Q_\theta$.