Learned JPEG Compression for DNN Vision

2026-06-15 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

Learned JPEG Compression for DNN Vision (J4D) is a novel training framework designed to optimize JPEG encoding parameters specifically for deep neural network (DNN) inference performance, rather than human perception. Recognizing that a significant portion of image data is consumed by AI, J4D addresses the challenge of representing the JPEG codec and compression rate in a differentiable, closed-form manner. It achieves this by integrating a differentiable soft quantizer based on a probabilistic quantization scheme, enabling both a differentiable JPEG proxy and analytical computation of the coded source's entropy, which closely estimates the actual compression rate. This setup allows the optimization problem to be solved using backpropagation. Experimental results demonstrate J4D's superior performance, achieving an accuracy increase of up to 11.60% at the same compression rate, or a compression rate reduction of up to 80.05% at the same accuracy, compared to default JPEG. J4D also shows promise for designing universal JPEG encoding parameters for various DNN architectures.

Key takeaway

For Machine Learning Engineers optimizing image data pipelines for deep neural networks, J4D offers a significant advancement over traditional JPEG. You should consider integrating this learned compression framework. It can achieve up to 11.60% higher model accuracy at existing compression rates. Alternatively, you can reduce your data storage and bandwidth needs by up to 80.05% without sacrificing performance. This approach also opens avenues for developing universal compression settings across diverse DNN architectures.

Key insights

J4D optimizes JPEG for DNNs using a differentiable codec and rate estimator, significantly boosting accuracy or reducing file size.

Principles

• JPEG optimization for DNNs requires a differentiable codec.
• Probabilistic quantization enables analytical rate estimation.
• DNN-centric compression can yield substantial efficiency gains.

Method

J4D trains JPEG encoding parameters by incorporating a differentiable soft quantizer and an information-theoretic rate estimator, allowing backpropagation to minimize rate while maximizing DNN inference performance.

In practice

• Apply J4D to improve DNN accuracy on compressed images.
• Reduce storage/bandwidth for DNN image datasets.
• Explore universal JPEG parameters for diverse DNNs.

Topics

• JPEG Compression
• Deep Neural Networks
• Computer Vision
• Probabilistic Quantization
• Inference Optimization
• Image Data Pipelines

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• What Matters in Practical Learned Image Compression — Optimizing learned image codecs for both perceptual quality and on-device speed yields superior compression and performance.
• Dual-Constrained Diffusion Image Compression for Operational Rate-Distortion-Perception Optimization — DCIC navigates the full RDP surface by integrating diffusion with dual distortion and idempotence constraints for flexible image compression.
• Prompt-Guided Image Editing with Masked Logit Nudging in Visual Autoregressive Models — Masked Logit Nudging enables efficient, high-quality prompt-guided image editing in visual autoregressive models.
• From Pixels to DNA: Why the Future of Compression Is About Every Kind of Data — Data compression is evolving into a foundational technology for all digital data, shifting from human-centric media to machine-centric semantics.
• Residual Modeling for High-Fidelity Learned Compression of Scientific Data — Tailored residual representations preserve learned compression advantages in high-fidelity scientific data scenarios.
• Exploiting Semantic and Pixel Representations for Ultra-Low Bitrate Image Compression — Ultra-low bitrate image compression can achieve pixel-level fidelity by combining semantic and pixel representations in a diffusion model.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.