QCNN-H: Single-Image Dehazing Using Quaternion Neural Networks

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Single-image haze removal is an ill-posed problem, making a universal solution difficult.
• Existing methods struggle with diverse real-world scenarios and applications.

Purpose of the Study:

• To develop a robust single-image dehazing method using a novel quaternion neural network.
• To evaluate the proposed method's performance in image dehazing and its impact on object detection.

Main Methods:

• A novel robust quaternion neural network architecture based on an encoder-decoder model was proposed.
• The network leverages quaternion image representation end-to-end with a quaternion pixel-wise loss function and quaternion instance normalization.
• The QCNN-H framework was evaluated on synthetic, real-world, and task-oriented datasets.

Main Results:

• The QCNN-H framework significantly outperformed state-of-the-art haze removal techniques in visual quality and quantitative metrics.
• Object detection accuracy and recall were improved in hazy scenes when using the QCNN-H method.
• This work represents the first application of quaternion convolutional networks to image dehazing.

Conclusions:

• The proposed QCNN-H method offers a robust and effective solution for single-image haze removal.
• Quaternion neural networks show promise for advancing image processing tasks, particularly in adverse conditions.
• The framework's ability to improve downstream tasks like object detection highlights its practical utility.