Remote sensing image denoising using residual encoder-decoder networks with edge enhancement

Remote sensing images are often affected by noise in the process of digitization and transmission processes. Denoising is an indispensable way of improving image quality. Despite showing an excellent noise removal performance, the existing denoising algorithms however typically suffer from a common drawback. Specifically, in the learning process, some edge information is lost, thereby over-smoothing the denoising result. Given the importance of details-including sharp edges and texture information-in remote sensing images, we propose a residual encoder-decoder denoising network with joint loss (REDJ) for GF-2 satellite data. Inspired by U-net, we use a deep convolutional framework is used to learn the end-to-end mapping from noisy images to the original ones. The encoder acts as a feature extractor that captures semantic information of image contents while eliminating noise, whereas the decoder recovers the image details. The high-resolution features from the encoder are combined with the up-sampled output by skip connection. We also introduce high-frequency decomposition and residual mapping to simplify the training process by reducing the solution space. As for the loss function, we modify the traditional denoising per-pixel loss. Given a well-trained convolutional neural network for defining perceptual loss, we instead to learn the perceptual differences of the extracted features instead of merely matching the low-level pixel information. Unlike the loss of detail resulting from normal per-pixel MSE loss, we recommend a new joint loss that combines the advantages of both per-pixel reconstruction and feature reconstruction, preserves additional edge and texture information, and generates clear denoised results. We employ the GF-2 satellite images in the experiments. To obtain enough training and testing data, we divide the entire high-resolution image is divided into 1200 pictures of size 512 and then allocate 70% of these images for training and the other 30% for testing. We generate the noisy images by adding Gaussian noise. To verify the effectiveness of our proposed network, we compare our quantitative and qualitative results with those of other state-of-the-art methods, including wavelet threshold, total variation, and K-SVD. Our proposed method REDJ can obtain the best index values both of PSNR and average gradient. In the qualitative visual sense, REDJ obtains clear denoising results because of the joint of perceptual loss. Compared with other methods that produce blurred regions generated by other methods, REDJ preserves more edge information and texture details. We also compare the run times of different methods for denoising images and find that REDJ has a relatively high CPU speed and achieves an excellent computational efficiency on GPU time. This paper successfully applies deep learning theory for denoising remote sensing images. We use the proposed network is used to remove noise from high-resolution GF-2 remote sensing images and to preserve the edge contours and fine details, which is conducive to facilitate later detection, classification, and other remote sensing applications. In our future work, we will explore to handle other types of noises, especially the complex real-world noises, and consider a single comprehensive network for more image restoration tasks. © 2020, Science Press. All right reserved.