Just Noticeable Difference Modeling for Deep Visual Features

2026-01-29

AI Summary

This work addresses the challenge of quantifying and controlling tolerable perturbations in deep visual features under resource-constrained settings to preserve downstream task performance. The authors propose FeatJND, the first method to introduce the concept of Just Noticeable Difference (JND) into deep feature space, establishing a task-aligned boundary for perturbation tolerance. By estimating normalized segmentation points, FeatJND computes a per-feature maximum tolerable perturbation map, enabling both feature importance visualization and token-level dynamic quantization. Experiments demonstrate that FeatJND-guided perturbations preserve task performance more effectively than Gaussian noise across image classification, object detection, and instance segmentation. Furthermore, under the same noise budget, its quantization strategy significantly outperforms random or globally uniform quantization baselines.

Technology Category

• Computer Vision: Adversarial Attacks & Robustness
• Machine Learning: Feature Construction/Reformulation
• Search and Optimization: Learning to Search

Application Category

• Responsible Web: Human-perceived consequences of algorithmic deployment on the web
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs

Abstract

Deep visual features are increasingly used as the interface in vision systems, motivating the need to describe feature characteristics and control feature quality for machine perception. Just noticeable difference (JND) characterizes the maximum imperceptible distortion for images under human or machine vision. Extending it to deep visual features naturally meets the above demand by providing a task-aligned tolerance boundary in feature space, offering a practical reference for controlling feature quality under constrained resources. We propose FeatJND, a task-aligned JND formulation that predicts the maximum tolerable per-feature perturbation map while preserving downstream task performance. We propose a FeatJND estimator at standardized split points and validate it across image classification, detection, and instance segmentation. Under matched distortion strength, FeatJND-based distortions consistently preserve higher task performance than unstructured Gaussian perturbations, and attribution visualizations suggest FeatJND can suppress non-critical feature regions. As an application, we further apply FeatJND to token-wise dynamic quantization and show that FeatJND-guided step-size allocation yields clear gains over random step-size permutation and global uniform step size under the same noise budget. Our code will be released after publication.

Problem

• Just Noticeable Difference
• Deep Visual Features
• Feature Quality
• Machine Perception
• Task Performance

Innovation

• Just Noticeable Difference
• Deep Visual Features
• Task-aligned Tolerance
• Feature Perturbation
• Dynamic Quantization