Recently, many deep-learning-based pan-sharpening methods have been proposed for generating high-quality pan-sharpened (PS) satellite images. These methods focused on various types of convolutional neural network (CNN) structures, which were trained by simply minimizing a spectral loss between network outputs and the corresponding high-resolution (HR) multi-spectral (MS) target images. However, owing to different sensor characteristics and acquisition times, HR panchromatic (PAN) and low-resolution MS image pairs tend to have large pixel misalignments, especially for moving objects in the images. Conventional CNNs trained with only the spectral loss with these satellite image data sets often produce PS images of low visual quality including double-edge artifacts along strong edges and ghosting artifacts on moving objects. In this letter, we propose a novel loss function, called a spectral-spatial structure (S3) loss, based on the correlation maps between MS targets and PAN inputs. Our proposed S3 loss can be very effectively used for pan-sharpening with various types of CNN structures, resulting in significant visual improvements on PS images with suppressed artifacts.

In this paper, we present a novel hardware-friendly super-resolution (SR) method based on a convolutional neural network (CNN) and its dedicated hardware (HW) on field programmable gate array (FPGA). Although CNN-based SR methods have shown very promising results for SR, their computational complexities are prohibitive for hardware implementation. To the best of our knowledge, we are the first to implement a real-time CNN-based SR HW that upscales 2K full high-definition video to 4K ultra high-definition (UHD) video at 60 frames per second (fps). In our dedicated CNN-based SR HW, low-resolution input frames are processed line-by-line, and the number of convolutional filter parameters is reduced significantly by incorporating depth-wise separable convolutions with a residual connection. Our CNN-based SR HW incorporates a cascade of 1D convolutions having large receptive fields along horizontal lines while keeping vertical receptive fields minimal, which allows us to save required line memory space in achieving comparable SR performance against full 2D convolution operations. For efficient HW implementation, we use a simple and effective quantization method with little peak signal-to-noise ratio (PSNR) degradation. Also, we propose a compression method to efficiently store intermediate feature map data to reduce the number of line memories used in HW. Our HW implementation on the FPGA generates 4K UHD frames of higher PSNR values at 60 fps and shows better visual quality, compared with conventional CNN-based SR methods that are trained and tested in software.

Super-resolution (SR) has become more vital, because of its capability to generate high-quality ultra-high definition (UHD) high-resolution (HR) images from low-resolution (LR) input images. Conventional SR methods entail high computational complexity, which makes them difficult to be implemented for up-scaling of full-high-definition input images into UHD-resolution images. Nevertheless, our previous super-interpolation (SI) method showed a good compromise between Peak-Signal-to-Noise Ratio (PSNR) performances and computational complexity. However, since SI only utilizes simple linear mappings, it may fail to precisely reconstruct HR patches with complex texture. In this paper, we present a novel SR method, which inherits the large-to-small patch conversion scheme from SI but uses global regression based on local linear mappings (GLM). Thus, our new SR method is called GLM-SI. In GLM-SI, each LR input patch is divided into 25 overlapped subpatches. Next, based on the local properties of these subpatches, 25 different local linear mappings are applied to the current LR input patch to generate 25 HR patch candidates, which are then regressed into one final HR patch using a global regressor. The local linear mappings are learned cluster-wise in our off-line training phase. The main contribution of this paper is as follows: Previously, linear-mapping-based conventional SR methods, including SI only used one simple yet coarse linear mapping to each patch to reconstruct its HR version. On the contrary, for each LR input patch, our GLM-SI is the first to apply a combination of multiple local linear mappings, where each local linear mapping is found according to local properties of the current LR patch. Therefore, it can better approximate nonlinear LR-to-HR mappings for HR patches with complex texture. Experiment results show that the proposed GLM-SI method outperforms most of the state-of-the-art methods, and shows comparable PSNR performance with much lower computational complexity when compared with a super-resolution method based on convolutional neural nets (SRCNN15). Compared with the previous SI method that is limited with a scale factor of 2, GLM-SI shows superior performance with average 0.79 dB higher in PSNR, and can be used for scale factors of 3 or higher.

With the advent of ultrahigh-definition (UHD) video services, super-resolution (SR) techniques are often required to generate high-resolution (HR) images from low-resolution (LR) images, such as HD images. To generate such HR images and a video of UHD resolutions in limited computing devices with hardware and software, low complex but excellent SR methods are particularly required. In this paper, we present a novel and fast SR method, called super-interpolation (SI), by unifying an interpolation step and a quality-enhancement step. The proposed SI method utilizes edge-orientation (EO)-based pre-learned kernels, which inherits the simplicity of interpolation and the quality enhancement of SR. It performs SR directly from the initial resolution of an input image to the target resolution of an up-scaled output image without requiring any intermediate interpolated image. The proposed SI method involves offline training and online up-scaling phases. In the offline training phase, training LR image patches are clustered based on their edge orientations into different EO classes for which class-dependent linear mapping functions are learned between training LR and HR image patches. In up-scaling phase, an HR output image patch for each LR input image patch is generated by applying an appropriate linear mapping function selected based on the EO of LR input image patch. Our proposed SI method is intensively compared with the ten state-of-the-art SR methods for common image sets and many HD/UHD images. The experimental results show that the SI method yields the smallest running time and requires relatively small hardware resources. It outperforms the six state-of-the-art methods in average (peak signal-to-noise ratio) PSNR/(structural similarity) SSIM, and exhibits competitive or somewhat lower PSNR/SSIM performance compared with the others.

Self-example-based super-resolution (SR) methods utilize internal dictionaries to reconstruct a high-resolution (HR) image from a single low-resolution (LR) input image. In general, a square-sized patch is used to find the LR-HR correspondences in the dictionaries. However, this may be a difficult issue because the LR input image and the dictionaries are of different scales. Inspired by this observation, we propose a novel self-example-based SR method, using context-dependent multi-shaped subpatches. Each LR input patch is segmented into multiple subpatches according to the context of the patch, enabling us to extract the better LR-HR correspondences. Our experimental results show that the proposed subpatch-based SR generates competitive high-quality HR images compared to state-of-the-art methods, with visually sharper edges that result in better visual quality.