# Adaptive Feature Selection for No-Reference Image Quality Assessment using Contrastive Mitigating Semantic Noise Sensitivity

**Authors:** Xudong Li, Timin Gao, Xiawu Zheng, Runze Hu, Jingyuan Zheng, Yunhang Shen, Ke Li, Yutao Liu, Pingyang Dai, Yan Zhang, Rongrong Ji  
**arXiv:** [2312.06158](https://arxiv.org/abs/2312.06158)  
**Date:** 2023-12-11  
**Summary page:** https://www.summarizepaper.com/en/paper/2312.06158v1/

## Key points

- Current NR-IQA methods rely on feature extraction in upstream backbone networks
- Not all extracted features are relevant and some may be harmful
- Proposed Quality-Aware Feature Matching IQA metric (QFM-IQM) uses contrastive learning to remove harmful features
- QFM-IQM compares image pairs with similar semantic features but varying quality scores
- Adaptive adjustment of upstream task's features mitigates negative impact of irrelevant or harmful features
- Distillation framework expands dataset and improves model's generalization ability
- Experimental results show that QFM-IQM outperforms state-of-the-art NR-IQA methods on eight standard datasets
- QFM-IQM achieves impressive Performance Linear Correlation Coefficient (PLCC) values compared to existing methods
- QFM-IQM selectively utilizes relevant features while eliminating harmful ones for accurate image quality assessment

## Summary

The current state-of-the-art No-Reference Image Quality Assessment (NR-IQA) methods rely on feature extraction in upstream backbone networks, assuming that all extracted features are relevant. However, recent research argues that not all features are beneficial and some may even be harmful, necessitating careful selection. It has been observed empirically that many image pairs with small feature spatial distances can have significantly different quality scores. To address this issue, the authors propose a novel Quality-Aware Feature Matching IQA metric (QFM-IQM) that employs contrastive learning to remove harmful features from the upstream task. The approach enhances the semantic noise distinguish capabilities of neural networks by comparing image pairs with similar semantic features but varying quality scores. By adaptively adjusting the upstream task's features through introducing disturbance, the proposed method effectively mitigates the negative impact of irrelevant or harmful features. Furthermore, a distillation framework is utilized to expand the dataset and improve the model's generalization ability. This framework enables the model to learn from both original images and their corresponding distorted versions, enhancing its ability to assess image quality accurately. Experimental results demonstrate that the proposed approach outperforms state-of-the-art NR-IQA methods on eight standard NR-IQA datasets. It achieves impressive Performance Linear Correlation Coefficient (PLCC) values of 0.932 (compared to 0.908 in TID2013) and 0.913 (compared to 0.894 in LIVEC). These results highlight the effectiveness of QFM-IQM in accurately assessing image quality by selectively utilizing relevant features while eliminating harmful ones. In conclusion, this research presents an innovative approach for No-Reference Image Quality Assessment by introducing contrastive learning and adaptive feature selection techniques.