Multi-Branch Network for Color Image Denoising Using Dilated Convolution and Attention Mechanisms.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Image denoising is a challenging ill-posed problem in computer vision.
• Convolutional neural network (CNN)-based methods show promise but struggle with complex image content.
• Simple networks often fail to recover aesthetically pleasing images.

Purpose of the Study:

• To propose an improved image denoising method using a multi-branch network.
• To enhance the recovery of aesthetically pleasing images from noisy inputs.
• To address the limitations of simple denoising networks in handling complex image content.

Main Methods:

• A multi-branch network based on an autoencoder architecture is proposed.
• The Pyramid Context Module (PCM) is integrated to enlarge the receptive field using dilated convolution.
• The Residual Bottleneck Attention Module (RBAM) is incorporated to refine features and reduce artifacts.

Main Results:

• The proposed network effectively learns multi-level contextual features.
• PCM successfully addresses the loss of global information.
• RBAM eliminates degraded features and minimizes undesired artifacts.
• Extensive experiments demonstrate superior performance over state-of-the-art methods.

Conclusions:

• The proposed multi-branch network significantly improves image denoising performance.
• The integration of PCM and RBAM modules enhances feature extraction and artifact reduction.
• The method achieves superior objective and subjective results compared to existing deep-learning approaches.