Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Image Quality Assessment: Investigating Causal Perceptual Effects with Abductive Counterfactual Inference

Wenhao Shen, Mingliang Zhou, Yu Chen, Xuekai Wei, Jun Luo, Huayan Pu, Weijia Jia

TL;DR

The paper reframes full-reference image quality assessment as a causal inference problem, introducing abductive counterfactual reasoning to uncover which deep features causally influence perceived quality. By decomposing deep representations into causally relevant components $\gamma$ and noise $\eta$ and employing interventions within a structural causal model, the method achieves backbone-independent perceptual quality scoring with robust generalization. It demonstrates competitive results across six IQA benchmarks and provides abductive validation showing improved interpretability of quality predictions. The framework offers a principled path to align IQA with human perception by isolating causal mechanisms, with potential applications beyond IQA to broader vision tasks.

Abstract

Existing full-reference image quality assessment (FR-IQA) methods often fail to capture the complex causal mechanisms that underlie human perceptual responses to image distortions, limiting their ability to generalize across diverse scenarios. In this paper, we propose an FR-IQA method based on abductive counterfactual inference to investigate the causal relationships between deep network features and perceptual distortions. First, we explore the causal effects of deep features on perception and integrate causal reasoning with feature comparison, constructing a model that effectively handles complex distortion types across different IQA scenarios. Second, the analysis of the perceptual causal correlations of our proposed method is independent of the backbone architecture and thus can be applied to a variety of deep networks. Through abductive counterfactual experiments, we validate the proposed causal relationships, confirming the model's superior perceptual relevance and interpretability of quality scores. The experimental results demonstrate the robustness and effectiveness of the method, providing competitive quality predictions across multiple benchmarks. The source code is available at https://anonymous.4open.science/r/DeepCausalQuality-25BC.

Image Quality Assessment: Investigating Causal Perceptual Effects with Abductive Counterfactual Inference

TL;DR

The paper reframes full-reference image quality assessment as a causal inference problem, introducing abductive counterfactual reasoning to uncover which deep features causally influence perceived quality. By decomposing deep representations into causally relevant components and noise and employing interventions within a structural causal model, the method achieves backbone-independent perceptual quality scoring with robust generalization. It demonstrates competitive results across six IQA benchmarks and provides abductive validation showing improved interpretability of quality predictions. The framework offers a principled path to align IQA with human perception by isolating causal mechanisms, with potential applications beyond IQA to broader vision tasks.

Abstract

Existing full-reference image quality assessment (FR-IQA) methods often fail to capture the complex causal mechanisms that underlie human perceptual responses to image distortions, limiting their ability to generalize across diverse scenarios. In this paper, we propose an FR-IQA method based on abductive counterfactual inference to investigate the causal relationships between deep network features and perceptual distortions. First, we explore the causal effects of deep features on perception and integrate causal reasoning with feature comparison, constructing a model that effectively handles complex distortion types across different IQA scenarios. Second, the analysis of the perceptual causal correlations of our proposed method is independent of the backbone architecture and thus can be applied to a variety of deep networks. Through abductive counterfactual experiments, we validate the proposed causal relationships, confirming the model's superior perceptual relevance and interpretability of quality scores. The experimental results demonstrate the robustness and effectiveness of the method, providing competitive quality predictions across multiple benchmarks. The source code is available at https://anonymous.4open.science/r/DeepCausalQuality-25BC.

Paper Structure

This paper contains 15 sections, 8 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Deep Network-based FR-IQA
  4. Causal Inference in Deep Learning
  5. Problem Formulation
  6. Methodology
  7. Framework
  8. Deep Causal Measurement
  9. Connection to Existing Methods and HVS
  10. Experimental Results
  11. Experimental Settings
  12. Quality Prediction Performance
  13. Ablation Analysis
  14. Qualitative Evaluation
  15. Conclusion

Figures (3)

  • Figure 1: The framework of our proposed method. Details of the confounder dictionary embedding and the structural causal model are detailed below.
  • Figure 2: Scatterplot of prediction results for different methods across databases. We performed normalization on both labels and predictions and 4-parameter logistic function is utilized before comparison, as suggested in VQEG, and the values of the lower-better methods are reversed before regression.
  • Figure 3: Effects of causal interventions on the perceptual representations of images under various distortions. The displayed channels, from top to bottom, correspond to the second, fourth, and fifth stages of the VGG-16 architecture.