Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LAR-IQA: A Lightweight, Accurate, and Robust No-Reference Image Quality Assessment Model

Nasim Jamshidi Avanaki, Abhijay Ghildyal, Nabajeet Barman, Saman Zadtootaghaj

TL;DR

A compact, lightweight NR-IQA model that achieves state-of-the-art (SOTA) performance on ECCV AIM UHD-IQA challenge validation and test datasets while being also nearly 5.7 times faster than the fastest SOTA model.

Abstract

Recent advancements in the field of No-Reference Image Quality Assessment (NR-IQA) using deep learning techniques demonstrate high performance across multiple open-source datasets. However, such models are typically very large and complex making them not so suitable for real-world deployment, especially on resource- and battery-constrained mobile devices. To address this limitation, we propose a compact, lightweight NR-IQA model that achieves state-of-the-art (SOTA) performance on ECCV AIM UHD-IQA challenge validation and test datasets while being also nearly 5.7 times faster than the fastest SOTA model. Our model features a dual-branch architecture, with each branch separately trained on synthetically and authentically distorted images which enhances the model's generalizability across different distortion types. To improve robustness under diverse real-world visual conditions, we additionally incorporate multiple color spaces during the training process. We also demonstrate the higher accuracy of recently proposed Kolmogorov-Arnold Networks (KANs) for final quality regression as compared to the conventional Multi-Layer Perceptrons (MLPs). Our evaluation considering various open-source datasets highlights the practical, high-accuracy, and robust performance of our proposed lightweight model. Code: https://github.com/nasimjamshidi/LAR-IQA.

LAR-IQA: A Lightweight, Accurate, and Robust No-Reference Image Quality Assessment Model

TL;DR

A compact, lightweight NR-IQA model that achieves state-of-the-art (SOTA) performance on ECCV AIM UHD-IQA challenge validation and test datasets while being also nearly 5.7 times faster than the fastest SOTA model.

Abstract

Recent advancements in the field of No-Reference Image Quality Assessment (NR-IQA) using deep learning techniques demonstrate high performance across multiple open-source datasets. However, such models are typically very large and complex making them not so suitable for real-world deployment, especially on resource- and battery-constrained mobile devices. To address this limitation, we propose a compact, lightweight NR-IQA model that achieves state-of-the-art (SOTA) performance on ECCV AIM UHD-IQA challenge validation and test datasets while being also nearly 5.7 times faster than the fastest SOTA model. Our model features a dual-branch architecture, with each branch separately trained on synthetically and authentically distorted images which enhances the model's generalizability across different distortion types. To improve robustness under diverse real-world visual conditions, we additionally incorporate multiple color spaces during the training process. We also demonstrate the higher accuracy of recently proposed Kolmogorov-Arnold Networks (KANs) for final quality regression as compared to the conventional Multi-Layer Perceptrons (MLPs). Our evaluation considering various open-source datasets highlights the practical, high-accuracy, and robust performance of our proposed lightweight model. Code: https://github.com/nasimjamshidi/LAR-IQA.
Paper Structure (25 sections, 3 equations, 2 figures, 8 tables)

This paper contains 25 sections, 3 equations, 2 figures, 8 tables.

Table of Contents

  1. Introduction
  2. Contributions
  3. Lower Complexity.
  4. Generalizability.
  5. Robustness and Accuracy.
  6. Related Work
  7. No-Reference Image Quality Assessment
  8. Generalization Across Diverse Datasets
  9. Kolmogorov-Arnold Networks (KAN)
  10. Proposed Method
  11. Datasets and Evaluation Metrics
  12. Data Pre-processing
  13. Quality Regression Module
  14. Image Encoder Module
  15. Loss Functions
  16. ...and 10 more sections

Figures (2)

  • Figure 1: Proposed model architecture. The image quality is evaluated using two branches: Authentic and Synthetic. In both branches, MobileNetV3 howard2019searching serves as the lightweight image encoder. The features extracted from these branches are concatenated and then used as input to KAN, the quality regression module, which outputs the final predicted image quality score.
  • Figure 2: MOS predictions of the Synthetic model using MobileNetV3, trained on KADID-10K and tested on sample images from the TID2013 dataset across different color spaces. Actual MOS scores are highlighted above in the pink colored bar. The change in predicted MOS scores before and after pre-training with color space loss ($\mathcal{L}_{rob}$) across various color spaces is shown in the blue colored bar. One can note that the predicted MOS scores for various images across different color spaces converge, indicating the robustness of the metric to different color spaces.