visual perception preserved denoising network in Image translation

Image denoising is a fundamental problem in computer vision and has received much attention from scholars. With the fast development of convolutional neural networks, more and more deep learning-based noise reduction algorithms have emerged. However, current image denoising networks tend to apply image noise reduction only in the RGB color space, ignoring the information at the visual perception level, making the images generated by these algorithms too smooth and lacking texture and details. Therefore, this paper proposes a novel noise reduction network in the image translation area using deep learning feature space instead of the traditional RGB color space to restore more realistic and more detailed texture information in generated images. The network contains a visual perception generator and a multi-objective optimization network. The generator includes a multiscale encoding-decoding sub-network, which extracts high-level perception features from input images. The optimization network contains content consistency loss, multiscale adversarial generation loss, and discriminator feature alignment loss, which effectively retains detailed texture information in the images. We synthesized noise of suitable intensity based on publicly available data sets and conducted multiple experiments to verify the effectiveness of the algorithm. The experimental results show that the proposed algorithm significantly improves textures and details in denoised images. The algorithm removes a large amount of noise information while preserving lots of perceptual information at the visual level, generating more realistic images with detailed texture features.