Hyun-Seung Lee

Evaluating super resolution algorithms

This study intends to establish a sound testing and evaluation methodology based upon the human visual characteristics for appreciating the image restoration accuracy; in addition to comparing the subjective results with predictions by some objective evaluation methods. In total, six different super resolution (SR) algorithms - such as iterative back-projection (IBP), robust SR, maximum a posteriori (MAP), projections onto convex sets (POCS), a non-uniform interpolation, and frequency domain approach - were selected. The performance comparison between the SR algorithms in terms of their restoration accuracy was carried out through both subjectively and objectively. The former methodology relies upon the paired comparison method that involves the simultaneous scaling of two stimuli with respect to image restoration accuracy. For the latter, both conventional image quality metrics and color difference methods are implemented. Consequently, POCS and a non-uniform interpolation outperformed the others for an ideal situation, while restoration based methods appear more accurate to the HR image in a real world case where any prior information about the blur kernel is remained unknown. However, the noise-added-image could not be restored successfully by any of those methods. The latest International Commission on Illumination (CIE) standard color difference equation CIEDE2000 was found to predict the subjective results accurately and outperformed conventional methods for evaluating the restoration accuracy of those SR algorithms.

Self-synthesis based super-resolution for a natural image

In this paper, we propose self-synthesis based super-resolution algorithm for natural images. The proposed method analyzes the property of the natural image and focuses on the self-similarity of natural objects. Structure and texture components are extracted from the self-image and synthesized on the interpolated image with matching weights to restore the self-similarity of interpolated image. Then, the synthesized image is compensated with MAP estimator to prevent artifacts. Experimental results show that the proposed method gives more natural texture in terms of human perception.

Simultaneous subtotal pancreatectomy and streptozotocin injection for diabetes modeling in cynomolgus monkeys

BACKGROUND In an experimental animal model of islet transplantation, stable induction of insulin-dependent diabetes mellitus (IDDM) and islet isolation from donor pancreas are essential. Total pancreatectomy for IDDM induction and islet procurement in nonhuman primates leads to unwanted loss of exocrine function and may lead to morbidities associated with IDDM. METHODS IDDM induction with streptozotocin (STZ) is associated with drug toxicity of STZ and necessitates the killing of another animal for islet procurement. In this study, we performed a subtotal pancreatectomy combined with reduced STZ injection to induce IDDM and procure islets in a nonhuman primate model. RESULTS Twelve cynomolgus monkeys received low-dose STZ injections (60 mg/kg) simultaneously with subtotal pancreatectomy. All monkeys recovered from the procedure without complications. IDDM was induced in the animals. 57,691 ± 16,050 islets were isolated from the resected pancreas and transplanted into other monkeys. CONCLUSIONS Simultaneous subtotal pancreatectomy and low-dose STZ injection represent an effective and safe method to create an animal model of insulin dependence diabetes, while at the same time providing sufficient amounts of fresh islet cells for allotransplantation without requiring killing of additional animals.

LPI adaptive descreening method with Hadamard domain analysis

This paper presents a new line-per-inch (LPI) adaptive descreening algorithm for restoring a high-quality image when scanning a printed matter. The LPI of scanned image is estimated and image region is classified into flat or edge region by analyzing the distribution in Hadamard transform coefficients. Then, an LPI adaptive edge preserving filtering is applied to reduce the halftone patterns not only in flat region but also in edge region. Proposed method has hardware-friendly structure without iteration and cascades of filters, and also it needs only integer adder and shift operations. Experimental result shows that the proposed method performs adaptively for a variety of LPI contents.

A new image super resolution by texture transfer

This paper proposes the new super resolution algorithm for improving the texture of low-quality image. Conventional methods have restored the high-resolution image based on the linear blur model with a variety of priors, which boost the frequency powers of the images while preventing from being noisy or having jagged edges. Previous methods show good performance for restoring the sharp and jagged-free edges. However, in the point of texture, they often fail to restore and result in cartoon-like images. To solve this problem, proposed method focuses on the property of natural textures and tries to create the fine texture from self-image by texture transfer, which modulates the phase of the high-frequency components of image in spatial and frequency domains. The sign of DCT coefficients are modulated to change the phase of image in frequency domain, and the high-frequency components of image are reshaped with an auto regressive filter in spatial domain. Moreover, the external noise sources are added to enhance the pixel-precision textures. Experimental results show that the texture of low-quality images is highly improved than that of previous methods. The fine textures created by proposed method give more natural feeling of highresolution image than that of results restored without a consideration of restoring textures.

Region-based printing quality enhancement

This paper presents a new region-adaptive printing quality enhancement algorithm for mixed scanning documents. To adapt a variety of line-per-inch (LPI) printing documents, novel pixel-wise LPI estimation is done via hadamard transform-based local frequency components analysis. Then, each pixel is classified into either flat-region or edge-region by comparing edge-region representative frequency peak power value and its corresponding flat-region representative frequency peak power value. After, LPI region adaptive-halftone-noise smoother and least square optimal restoration filter enhance the quality of scanned documents. Experimental results indicate that the proposed one provides outperforms other state of the art noise reduction methods on a variety of LPI documents.

Fine grain creation for UHDTV

A novel fine grain creation method for Ultra-HighDefinition TeleVision (UHDTV) has been proposed. To make fine grain components, multi-layer super resolution reconstruction and the frequency modulation method are developed. Base and Texture layer of images are fused separately to prevent the loss of texture components in interpolation. After, fine grains are created on up-scaled images. Fine grain creation is composed of two parts, one is diffusion of low frequency components, and the other is generation of aliased high frequency components. Those are added on the up-scaled video with matching weights to enhance the feeling of textures. The proposed method shows a better performance compared to the conventional restoration methods in terms of texture by the experimental results.

Edge enhancement of a sub-pixel rendered image for printing system

Recently, sub-pixel rendered filter is often used as a post processing of graphic contents to reduce the jagging artifacts in displays. However, in a printing system, sub-pixel filtered image yields blurring and jagging artifacts due to the difference of displaying method. In order to prevent artifacts, new edge enhancement method is proposed. Edge region of an image is processed by an adaptive intensity mapping function to reduce the mid-range of intensities that make jagging artifacts in half-toning. Then, a color replacement function is applied to reduce the color fringe artifacts. Experimental results show that the proposed algorithm improves the quality of printed sub-pixel rendered image.