A Generative Adversarial Network-Based Image Denoiser Controlling Heterogeneous Losses.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Image noise significantly degrades visual information quality.
• Existing denoising methods, such as convolutional neural networks (CNNs), often struggle with preserving structural details.
• Generative Adversarial Networks (GANs) show promise but require novel loss functions for enhanced performance.

Purpose of the Study:

• To develop an advanced image denoising technique using a novel GAN architecture.
• To enhance the restoration of structural information in noisy images.
• To reduce computational complexity while improving denoising performance.

Main Methods:

• A novel generative adversarial network (GAN) framework for image denoising.
• Implementation of heterogeneous losses, including structural loss and mean squared error (MSE) loss.
• Adaptive adjustment of structural loss strength by the discriminator for each input patch.
• Integration of a depthwise separable convolution module with dilated convolution and symmetric skip connections.

Main Results:

• The proposed GAN method demonstrated superior performance compared to existing CNN denoisers.
• Significant improvements in visual information fidelity (up to 0.027) and feature similarity index (up to 0.008) were achieved.
• The depthwise separable convolution module reduced computational complexity.

Conclusions:

• The novel GAN-based image denoising method effectively preserves structural information and enhances image quality.
• Adaptive heterogeneous loss functions are crucial for improving GAN-based image restoration.
• The proposed architecture offers a computationally efficient and high-performing solution for image denoising.