The Last Byte: Learning Just Enough for Machine-Oriented Image Compression

Wuyuan Xie, Zhenming Li, Ye Liu, Jian Jin, Yun Song, Miaohui WangAAAI 2026 · DOI

Abstract

Just recognizable distortion (JRD) has been introduced for image compression for machines, aiming to quantify the maximum coding distortion that can be tolerated by a specific perception model, thereby defining the upper bound of machine vision redundancy (MVR). However, existing JRD-based redundancy estimation methods face three key challenges: limited dataset annotation accuracy, low prediction efficiency, and insufficient perception accuracy, all of which hinder their practical deployment. To address these limitations, we propose a new MVR-Net, a frame-wise efficient JRD prediction method that generates the optimal encoding quantization map in a single inference pass. Furthermore, we refine the annotation standard for JRD datasets based on experimental insights, enhancing the precision of recognizable redundancy measurement. Compared to state-of-the-art methods, MVR-Net achieves a superior balance between bitrate reduction and perception accuracy in JRD-guided compression, while offering up to a 40,000× speed improvement, demonstrating its practicality and efficiency for real-world applications.

Introduction

Traditional image compression standards, such as JPEG or H.265/HEVC, are optimized for the Human Visual System (HVS), discarding data imperceptible to the human eye. However, as autonomous systems and AI-driven monitoring become ubiquitous, images are increasingly consumed by machines rather than humans. Machines "see" differently; they may be sensitive to pixel-level variations ignored by humans while remaining robust to significant distortions in non-essential background regions.

The threshold of maximum tolerable distortion for a machine vision model is defined as Just Recognizable Distortion (JRD). Current JRD estimation methods are often inaccurate due to coarse labeling, inefficient because of iterative encoding loops, and difficult to scale across complex scenes. This paper introduces MVRNet, a framework designed to identify the "last byte" of data required for machine accuracy, achieving significant speedups and bitrate reductions.

Methods: The MVRNet Architecture

MVRNet is a "no-reference," frame-wise model that predicts an image's redundancy in a single pass. Unlike previous methods that require the original image for comparison or multiple inference loops, MVRNet processes the entire frame at once.

The MVRNet pipeline for JRD-guided machine-oriented image compression.

The architecture utilizes a dual-path approach:

  1. The Task Path: Processes a Confidence Map generated by a pre-trained object detector. This map serves as "perceptual knowledge," indicating regions where the AI model is already certain.
  2. The Pixel Path: Analyzes raw pixels to capture low-level visual cues like edges and textures.

These paths are integrated via the Efficient Feature Extraction Module (EFEM) and the Efficient Feature Fusion Module (EFFM). The output is a Quantization Parameter (QP) map that guides the video encoder (e.g., VVC) on how much to compress specific blocks of the image.

Dataset: MVRSet

To train MVRNet, the researchers developed MVRSet, derived from 10,000 images in the COCO2017 dataset. They generated 640,000 distorted versions using the VVC (H.266) standard across 64 quality levels. Crucially, MVRSet includes precise annotations for both foreground objects and background regions, enabling the model to learn that backgrounds can often be compressed more aggressively without impacting task performance.

Results and Discussion

MVRNet demonstrates a paradigm shift in efficiency. It processes a frame in approximately 0.17 seconds, a 41,859× improvement over previous state-of-the-art methods like OWJRD, which took over 7,000 seconds per frame.

Visualization of JRD QP maps, showing region-specific compression levels.

A key insight is that detection confidence scores serve as a robust proxy for JRD. High confidence in the original image allows for higher compression on that object before accuracy degrades. On the COCO dataset, MVRNet achieved a 0.8640 mAP at a bitrate of 0.5314 BPP. In contrast, standard VVC required a bitrate of 5.8720—over ten times more data—to achieve a lower accuracy of 0.7290.

Conclusion

MVRNet provides a practical, ultra-fast solution for machine-oriented image compression. While highly effective, future work could reduce dependency on specific detector architectures and expand the framework to tasks beyond object detection, such as action recognition or 3D analysis. Integrating these predicted QP maps directly into hardware-level VVC encoders remains a promising avenue for real-time streaming applications.

Summary updated

Abstract

Just recognizable distortion (JRD) has been introduced for image compression for machines, aiming to quantify the maximum coding distortion that can be tolerated by a specific perception model, thereby defining the upper bound of machine vision redundancy (MVR). However, existing JRD-based redundancy estimation methods face three key challenges: limited dataset annotation accuracy, low prediction efficiency, and insufficient perception accuracy, all of which hinder their practical deployment. To address these limitations, we propose a new MVR-Net, a frame-wise efficient JRD prediction method that generates the optimal encoding quantization map in a single inference pass. Furthermore, we refine the annotation standard for JRD datasets based on experimental insights, enhancing the precision of recognizable redundancy measurement. Compared to stateof-the-art methods, MVR-Net achieves a superior balance between bitrate reduction and perception accuracy in JRD-guided compression, while offering up to a 40,000× speed improvement, demonstrating its practicality and efficiency for real-world applications.
BibTeX
@inproceedings{xie2026last,
  author = {Wuyuan Xie and Zhenming Li and Ye Liu and Jian Jin and Yun Song and Miaohui Wang},
  title = {{The Last Byte: Learning Just Enough for Machine-Oriented Image Compression}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2026},
  doi = {10.1609/aaai.v40i19.38635},
}
Figure