CHEM: Estimating and Understanding Hallucinations in Deep Learning for Image Processing

2025-12-10

AI Summary

U-shaped reconstruction models (e.g., U-Net) commonly produce visually plausible but physically inconsistent hallucination artifacts in image deconvolution, compromising reliability in safety-critical applications. To address this, we propose the first distribution-agnostic and architecture-agnostic framework for hallucination quantification and attribution. Our method innovatively integrates wavelet/shearlet-based multi-scale feature extraction with conformal quantile regression to enable pixel-level hallucination localization and calibrated uncertainty estimation. Crucially, we provide the first theoretical analysis of hallucination origins in U-shaped networks from an approximation-theoretic perspective. The framework is compatible with diverse architectures—including U-Net, SwinUNet, and Learnlets—and demonstrates significant improvements in hallucination detection accuracy and trustworthiness assessment on the CANDELS astronomical dataset. This work establishes both theoretical foundations and practical tools for enhancing interpretability and robustness of reconstruction models in high-stakes domains.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Computer Vision: Adversarial Attacks & Robustness
• Reasoning under Uncertainty: Uncertainty Representations

Application Category

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

Abstract

U-Net and other U-shaped architectures have achieved significant success in image deconvolution tasks. However, challenges have emerged, as these methods might generate unrealistic artifacts or hallucinations, which can interfere with analysis in safety-critical scenarios. This paper introduces a novel approach for quantifying and comprehending hallucination artifacts to ensure trustworthy computer vision models. Our method, termed the Conformal Hallucination Estimation Metric (CHEM), is applicable to any image reconstruction model, enabling efficient identification and quantification of hallucination artifacts. It offers two key advantages: it leverages wavelet and shearlet representations to efficiently extract hallucinations of image features and uses conformalized quantile regression to assess hallucination levels in a distribution-free manner. Furthermore, from an approximation theoretical perspective, we explore the reasons why U-shaped networks are prone to hallucinations. We test the proposed approach on the CANDELS astronomical image dataset with models such as U-Net, SwinUNet, and Learnlets, and provide new perspectives on hallucination from different aspects in deep learning-based image processing.

Problem

• Quantifying hallucinations in deep learning image reconstruction
• Understanding why U-shaped networks generate unrealistic artifacts
• Ensuring trustworthy computer vision models in safety-critical scenarios

Innovation

• Wavelet and shearlet representations extract hallucination features
• Conformalized quantile regression assesses hallucination levels distribution-free
• Method applicable to any image reconstruction model for quantification