A New Edge and Pixel-Based Image Quality Assessment Metric for Colour and Depth Images

Author: Seyed Muhammad Hossein Mousavi
Contact: mosavi.a.i.buali@gmail.com

Please cite

• Mousavi, Seyed Muhammad Hossein, and S. Muhammad Hassan Mosavi. "A new edge and pixel-based image quality assessment metric for colour and depth images." 2022 9th Iranian Joint Congress on Fuzzy and Intelligent Systems (CFIS). IEEE, 2022.

Link to the paper:

• https://ieeexplore.ieee.org/document/9756490
• DOI: 10.1109/CFIS54774.2022.9756490

Link to the NDDB dataset:

• https://www.kaggle.com/datasets/hosseinmousavi/noisy-depth-database-nddb

Overview

This repository is dedicated to the paper "A New Edge and Pixel-Based Image Quality Assessment Metric for Colour and Depth Images", which introduces a novel Full-Reference Image Quality Assessment (FR-IQA) metric called EPIQA. The method evaluates image quality by combining edge-based and pixel-based features, delivering enhanced performance compared to traditional IQA metrics.

---

Abstract

Measuring the quality of digital images is crucial in image processing, especially for color and depth images. This paper proposes EPIQA (Edge and Pixel-based Image Quality Assessment), a new FR-IQA metric that combines improved edge-based IQA methods with Peak Signal-to-Noise Ratio (PSNR). EPIQA addresses the weaknesses of traditional metrics like MSE, PSNR, and SSIM by leveraging the edge and pixel-level features of images.

Key Contributions:

• Introduced EPIQA, which combines edge-based features and pixel-level comparisons.
• Validated the method on color and depth image databases, including the creation of a Noisy Depth Database (NDDB).
• Demonstrated superior performance over existing IQA metrics using standard benchmarks like SROCC, KROCC, PLCC, and RMSE.

---

Methodology

The proposed EPIQA metric works as follows:

1. Preprocessing:

   • Median filtering to remove noise while preserving edges.
   • Unsharp masking to enhance edge clarity.

2. Feature Extraction:

   • The image is divided into 8x8 blocks.
   • Five edge-based features are extracted:
     • Edge Density (ED): Proportion of edge pixels.
     • Edge Length Average (ELA): Average length of detected edges.
     • Gray Level Region (GLR): Number of unique intensity levels.
     • Number of Edge Pixels (NEP): Total edge pixels per block.
     • Edge Orientation (EO): Count of vertical and horizontal edges.

3. Metric Calculation:

   • Compute the Euclidean distance between the feature vectors of reference and distorted images.
   • Combine the normalized distance with the PSNR to generate the final EPIQA score.

4. Validation:

   • EPIQA is validated using four performance metrics:
     • Spearman Rank-Order Correlation Coefficient (SROCC)
     • Kendall Rank-Order Correlation Coefficient (KROCC)
     • Pearson Linear Correlation Coefficient (PLCC)
     • Root-Mean-Square Error (RMSE)

---

Results

Performance Comparison

EPIQA was tested on various benchmark databases and a newly created Noisy Depth Database (NDDB), achieving the following:

1. Databases:

   • A57 Database (color images with distortions)
   • TID2008 Database (color images with 68 distortion types)
   • Eurecom Kinect Database (depth images)
   • Proposed NDDB (depth images with 7 types of manually added noise)

2. Metrics:

   • EPIQA consistently outperformed traditional IQA metrics like MSE, PSNR, and SSIM.
   • Demonstrated high correlation with subjective evaluations (MOS).

Validation Metrics (Sample Results):

Database	SROCC	KROCC	PLCC	RMSE
A57	0.874	0.666	0.901	0.102
TID2008	0.916	0.777	0.823	0.106
Eurecom Kinect	0.742	0.775	0.781	0.106
NDDB	0.783	0.750	0.748	0.101

---

Proposed Noisy Depth Database (NDDB)

To address the lack of proper depth image databases, this paper introduces the Noisy Depth Database (NDDB):

• Contains 30 depth images of small home objects.
• Includes 7 types of noise:
  • Gaussian, Salt-and-Pepper, Poisson, Speckle, Quantization, Additive White Gaussian, Blockwise.
• Provides a benchmark for testing IQA metrics on depth images.

---

Applications

The proposed EPIQA metric has applications in:

• Quality Control Systems: Assessing the quality of images in real-time systems.
• Image Restoration: Evaluating the effectiveness of denoising and enhancement algorithms.
• Compression Algorithms: Comparing the quality of compressed vs. original images.
• Depth Image Analysis: Extending IQA techniques to depth images in robotics, augmented reality, and more.

---

Future Work

1. Extend EPIQA to work with Reduced-Reference (RR) and No-Reference (NR) IQA systems.
2. Develop a more comprehensive database for depth and color images with diverse distortion types.
3. Incorporate context-based features to further enhance the metric's robustness.

---

• 10 IQA metrics would be run on single image, which are:
• (1) The Peak Signal to Noise Ratio (PSNR)
• (2) The Signal to Noise Ratio (SNR)
• (3) The Mean-Squared Error (MSE)
• (4) The R-mean-squared error (RMSE)
• (5) The Measure of Enhancement or Enhancement (EME)
• (6) The Structural Similarity (SSIM)
• (7) The Edge-strength Structural Similarity (ESSIM)
• (8) The Non-Shift Edge Based Ratio (NSER)
• (9) The Edge Based Image Quality Assessment (EBIQA)
• (10)The Edge and Pixel-Based Image Quality Assessment (EPIQA)