Gaussian noise removal with attention-based convolutional neural networks

The process of noise reduction in digital images involves mitigating distortions that occur, aiming to eliminate them while preserving the original structure of the image as much as possible. This operation holds significant importance as a preprocessing step in various image-processing studies due to its ability to enhance the visual quality of an image. Gaussian noise is one of the most encountered types of noise. In the scope of this thesis, a new network based on Convolutional Neural Networks (CNN) is proposed for the removal of Gaussian noise at different levels in grayscale and color images. Initially, a Local Binary Pattern (LBP) based texture image is extracted from the noisy image to serve as a guide for the main noise reduction process, aiming to preserve the original structure of the image. For this purpose, a CNN named TENet, with a two-branch structure, is proposed. Subsequently, for the noise reduction process, a new CNN-based MLFAN network is proposed, taking both the noisy image and the LBP-based texture image obtained from the noisy image as inputs. The MLFAN network is designed with convolutional layers having different kernel sizes for multi-level feature extraction and channel-based attention mechanisms for weighting these features. Additionally, a comprehensive ablation study is presented to evaluate the design of the MLFAN network, accompanied by numerical and visual results.