Shotaro Moriya

Advanced demosaicing methods based on the changes of colors in a local region

A large number of demosaicing methods using a directional or adaptive color correlation have been suggested. In such conventional methods, however, there is the problem of artifacts occurring on the edges of color boundaries. In the advanced demosaicing method we propose, demosaicing is carried out on the basis of the results obtained from the analysis of the changes of color signals in a local region. Using image simulation, we confirmed the artifact noise generated in the conventional method had been reduced by the advanced demosaicing method.

Resolution enhancement based on laplacian pyramid

Increasing pixel resolutions on displays are a current trend of consumer products. Lower resolution images can be input to these kinds of products. In this case, we have to enlarge the input images. Usually, large scale integration (LSI) and an application specified chip (ASIC) for consumer products enlarge input images by linear processes. However a linear process cannot generate a higher frequency component than the Nyquist Frequency of the input image and outputs a blurred image. In this paper, we present an image enlargement method that generates the higher frequency component and is suitable for hardware implementation.

Light-Weight Single Image Super-Resolution via Pattern-wise Regression Function.

We propose a novel upsampling approach that is suitable for hardware implementation. Compared with past super-resolution (SR) upsampling methods (e.g. example based upsampling), structure of our upsampling approach is very simple. Strategy of our approach is mainly 2 terms; off-line training term and realtime upscaling term. (i)During training term, grouping lowresolution (LR) high-resolution (HR) patch pairs and determined a linear regression function of each groups. And (ii)during upscaling term, assigning pattern number to each of input LR patches according to the signature using a local binary pattern (LBP), and transforming input LR patches to HR patches by applying the trained regression function based on the LBP in a patch-by-patch fashion. Our evaluation result shows that our method is comparable to other state-of-the-art methods. Furthermore, our approach is compactly implemented on LSI (e.g. FPGAs) or be shorten the processing time on software because of simplicity of the structure. Introduction Currently, security systems become an increased center of focus in our daily life. There is a demand for recognizing some distant objects such as faces or license plates on a big display. In order to do that, we upscale low-resolution (LR) images to high resolution (HR) images. And the transformation poses a problem because the new image should have more information of pixels than the original image. A simple method of increasing the image resolution is interpolation technique, such as bilinear and bicubic interpolation. Although these methods are quite simple to implement, these approaches cannot estimate and reconstruct high-frequency information in generated HR image, because these methods assumed image smoothness. Super-Resolution is another technique of upsampling method which generates an HR image from multiple LR images [1] or single LR image [2][3][4]. This upsampling is always ill-posed problem because amount of pixels in target image is typically higher than those in observed input image(s). Conventional SR methods use multiple images with small motion as inputs. Although those methods need a mechanics of moving camera system(e.g. aerial images), this approach can reconstruct latent high-frequency elements. Another type of SR technique uses database constructed by training. This approach typically learns a relationship between the HR patches and the downsampled LR patches in advance. The example based upsampling method [2][3] is one of those approach and it can estimate high-frequency information. It is known that the example based method has two problems. One is that this method should have a large amount of database in order to deal with large variation of real world scenes. This large variation of database requires huge computational resources and it is not suitable for embedded systems. The other problem is that resulting images sometimes include noise, halos, ringing, and aliasing artifacts, because of mismatch between the input image and the image database. Our goal is to generate high-resolution, artifact-less output images using reasonable computational resources. In Section 2 and 3, we describe our proposed method. In Section 4 we demonstrate the experimental results of our method compared with several SR methods. And finally in Section 5 we conclude the paper. Super Resolution via Pattern-wise Regression Function In this section, we describe our super-resolution algorithm. Our method generates a HR image from a LR image using linear regression functions. The method uses 2 main stages; off-line training term and real-time upscaling term. Data Training During training term (Fig.1.), our approach groups LR-HR patch pairs of similar pattern according to the signature using a local binary pattern (LBP). In each of the group, we determine a linear regression function that will estimate an output HR patch corresponding to the LR patch. First, LR image is obtained by downsamping the HR image. This LR image is used for discriminating LBP, which can determine the signature of each local area of input image (explained in latter section). Next, in order to group LR-HR patch, we upscale the downsampled LR image and obtain the same size, but low resolution compared with HR image. Then upscaled LR patch and corresponding HR patch are extracted (in this case, size of a patch is 5× 5 pixels). To eliminate the influence of a luminance offset of image patch, texture component and nontexture component (average luminance of a patch) are separated from the patch and texture components of LR and HR patches are used for making patch-pair. LR-HR patch pairs of all pixels are categorized into 128 groups on the basis of pattern number Pnum. To approximate a relationship between LR patch and HR patch, we determine a linear regression function that will estimate an output HR patch. Coefficient data M (25×25) is calculated according to below regression function equation. Then, Xt / Yt indicates a matrix of accumulated LR / HR texture patches in 1 LBP group. eye is an unit matrix and λ is a static parameter. LR patch. ©2016 Society for Imaging Science and Technology DOI: 10.2352/ISSN.2470-1173.2016.18.DPMI-028 IS&T International Symposium on Electronic Imaging 2016 Digital Photography and Mobile Imaging XII DPMI-028.1