Multi-scale wavelet and deep learning integrated adaptive weighted non-local means denoising algorithm

Image denoising is a critical component in numerous disciplines, including medical imaging, remote sensing, and computer vision. This paper presents an adaptive weighted non-local means (MAW-NLM) algorithm, which employs multi-scale wavelet decomposition and deep learning to address the shortcomings of conventional non-local means in more intricate scenarios. The initial step in the image decomposition process involves the division of the image into low- and high-frequency components through the implementation of multi-scale wavelet decomposition. This decomposition is followed by the processing of each component individually, with the objective of preserving edges and fine details with greater efficacy. A deep learning-based similarity metric, leveraging a convolutional neural network for high-dimensional feature extraction, replaces the conventional Euclidean distance, thus enhancing the robustness of the similarity measurement. Furthermore, the algorithm dynamically adjusts smoothing parameters based on local gradient and texture complexity, significantly improving its adaptability. The integration of a fast nearest-neighbor search algorithm and a two-stage optimization approach has been demonstrated to enhance both computational efficiency and denoising effectiveness. The experimental results demonstrate the superiority of MAW-NLM in comparison to traditional methods.