Noise reduction in digital images based on original RAW files using neural networks

This paper presents a noise reduction method for RAW photo images, focusing on preserving the original information and improving the processing quality. Digital image processing is important for surveillance and vision systems where quality and detail play a key role. The proposed method is based on a combination of UNet and HQSNet neural networks. HQSNet performs semi-square partitioning of the input data, emphasizing key regions and reducing the dimensionality of less significant ones. UNet, in turn, efficiently processes the prepared data, preserving high granularity and tone transitions. The method is tested on real images, including complex night portrait and starry sky scenes, demonstrating high performance on MSE metrics and expert evaluations. Comparison with traditional methods, such as median and Bilateral filters, showed the superiority of the new approach in both noise removal quality and image detail preservation. The advantages of the method include preservation of dynamic range and the possibility of deep post-processing. The results obtained confirm its effectiveness in digital processing problems, which makes the development promising for application in automatic image analysis and enhancement systems.