A study on multi-scale attention dense U-Net for image denoising method

Although many models that have applied learning exhibit good performance, the dataset or image generation and transmission process used for learning may contain noise, which cannot produce the expected results and performance. The representative image denoising technique using deep neural networks generates noisy images by forcibly adding special noise to the original image and learning to make it the same as the original image. However, the performance of deep neural networks depends on depth, and to improve performance, increasing only depth will reach a performance saturation state, which will encounter difficulties. In order to improve these issues, this article applies the Multi-scale Attention model to the representative denoising deep learning model U-Net, to suppress unnecessary information and provide functionality that only emphasizes important information. In a new modular approach, the given input value is divided into two parts based on its internal relationship: the part where the important parts are concentrated and the part where the important parts are concentrated through spatial information. The attention unit based Outburst structure, which combines the two parts after parallel execution, has been implemented, demonstrating better performance than existing models. Moreover, without adding too many parameters, more spatial feature maps than other models are generated by focusing on the effects of components, not only through PSNR and SSIM. The improved performance was also confirmed by removing noisy in images.