Dual-Stream Complex-Valued Convolutional Network for Authentic Dehazed Image Quality Assessment.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Evaluating the perceptual quality of dehazed images is a significant challenge in image processing.
• Existing no-reference image quality assessment methods often struggle with the complexities of dehazed imagery.

Purpose of the Study:

• To propose a novel no-reference complex-valued convolutional neural network (CV-CNN) model for automatic dehazed image quality evaluation.
• To enhance perceptual feature learning and generalization capabilities using complex-valued representations.

Main Methods:

• Development of a dual-stream CV-CNN architecture for dehazed image quality assessment (DQA).
• One stream processes the dehazed RGB image for distortion artifacts (distortion-sensitive stream).
• A second stream analyzes a novel dark channel difference image to detect residual haze (haze-aware stream).

Main Results:

• The proposed CV-CNN model demonstrates superior generalization capability in perceptual feature learning compared to real-valued networks.
• Experimental results on three public DQA databases validate the model's effectiveness.
• The CV-CNN DQA model outperforms existing state-of-the-art no-reference image quality assessment algorithms.

Conclusions:

• The novel dual-stream CV-CNN effectively evaluates the perceptual quality of dehazed images.
• Complex-valued representations offer advantages for learning discriminative features in DQA.
• The proposed model provides a robust and generalizable solution for no-reference DQA.