Related Experiment Video
Updated: Sep 9, 2025
07:34
Excitation-Scanning Hyperspectral Imaging Microscopy to Efficiently Discriminate Fluorescence Signals
Published on: August 22, 2019
8.1K
A Style Transfer-Based Fast Image Quality Assessment Method for Image Sensors
Weizhi Xian1,2, Bin Chen1,3, Jielu Yan4
1Chongqing Research Institute of Harbin Institute of Technology, Harbin Institute of Technology, Chongqing 401151, China.
Sensors (Basel, Switzerland)
|August 28, 2025
Summary
A new method for image quality assessment, Style- and Content-based IQA (SCIQA), uses deep learning and style transfer principles. SCIQA outperforms existing methods in accuracy and generalizability for evaluating visual data fidelity.
Area of Science:
- Computer Vision
- Image Processing
- Machine Learning
Background:
- Accurate image quality evaluation is crucial for optimizing sensor performance and visual data fidelity.
- Image style, encompassing visual characteristics like color and texture, significantly impacts perceived quality.
Purpose of the Study:
- To propose a novel full-reference image quality assessment (FR-IQA) method named Style- and Content-based IQA (SCIQA).
- To develop a method that integrates image content and style for more accurate perceptual quality assessment.
Main Methods:
- Utilized a deep convolutional neural network (CNN) for image feature decomposition in the deep domain.
- Defined a deep perceptual distance metric combining content-based and style-based measures.
- Formulated a perceptual optimization problem solved via convex optimization for model parameter tuning.
Main Results:
- SCIQA demonstrated superior performance over state-of-the-art FR-IQA methods on multiple benchmark datasets (LIVE, CSIQ, TID2013, KADID-10k, PIPAL).
- Achieved high Pearson linear correlation coefficients (PLCC) (e.g., 0.956 on LIVE) and robust generalizability (SROCC of 0.702 on PIPAL).
- Outperformed traditional (SSIM) and deep learning-based (DISTS) methods in accuracy.
Conclusions:
- SCIQA offers exceptional prediction accuracy, strong interpretability, and low computational complexity.
- The proposed method is a practical and effective tool for real-world image quality assessment applications.
