RYF-Net: Deep Fusion Network for Single Image Haze Removal.

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Single image dehazing is a complex problem, crucial for improving visual quality and downstream tasks.
• Accurate estimation of the scene transmission map (TrMap) is fundamental for effective haze removal.

Purpose of the Study:

• To develop a robust convolutional neural network (CNN) architecture for estimating accurate scene transmission maps (TrMap) from hazy images.
• To enhance the performance of single image dehazing and improve object detection in hazy environments.

Main Methods:

• A novel CNN architecture employing RNet and YNet to extract features from RGB and YCbCr color spaces, generating two TrMaps.
• A TrMap fusion network (FNet) to integrate the generated TrMaps for a more robust estimation.
• Evaluation using structural similarity index (SSIM), mean square error (MSE), and peak signal to noise ratio (PSNR) on multiple datasets.

Main Results:

• The proposed approach demonstrates superior performance compared to existing state-of-the-art methods in single image dehazing.
• Experiments conducted on diverse datasets (D-HAZY, ImageNet, Indoor SOTS, HazeRD, and real-world images) validate the method's effectiveness.
• Significant improvements were observed in object detection accuracy when using the proposed dehazing approach.

Conclusions:

• The developed CNN-based architecture effectively estimates robust scene transmission maps for single image dehazing.
• The proposed method offers a significant advancement in both image dehazing quality and the accuracy of high-level vision tasks like object detection in hazy scenes.