A Method of Image Denoising via Dense Attention DnCNN

Many models with deep learning exhibit good performance, but due to the noise generated during the dataset or image generation and transmission used for learning, it may include noise and cannot achieve the desired results and performance. The representative denoising technique of applying deep neural networks, DnCNN, forcibly generates noisy images by adding Gaussian noise to the original image and learning to be the same as the original image. Applying deep neural network denoising techniques to perform learning and denoising in this way. However, when performance depends on depth and simply increases depth to improve performance, deep neural networks will experience performance saturation. To improve this issue, this paper proposed an image processing system with a attention mechanism. This method improves the image processing method by applying the attention unit to the representative denoising deep learning model DnCNN, and converting it into an outburst structure. It is divided into two parts: focusing on internal relationships based on given input values and focusing on spatial information. These two parts are executed in a parallel structure and then combined. Focusing attention mechanism identifies and compresses important features, and quickly adding these features by highlighting the structure can effectively eliminate noise. In addition, by applying it to image processing methods that require noise removal, various applications can be carried out.