Abstract

A new learning-based single image super-resolution technique that upscales the low resolution (LR) image in a single pass toits desired high resolution (HR) image is proposed here.Inthe upscaling procedure, a linearmapping function is learned from the external data set. Mapping function converts LR patch to its corresponding patch.In most of the patch-based learning technique, smoothness of the overlapped regions is performed with an average value of the overlapped regions. As a result, edge information that reflects in adjacent LR patches does not always transparently reflects in HR patches. So in our technique, we applied edge directed smoothness in adjacent patches. An edge exists along the direction, where the second-order derivative is lower. To reach this, we have selected non-overlapping patch, and after getting HR patch, we performed edge directed smoothness of adjacent patches. This results smoothness of adjacent patches with more detailedge information. Apart from this nonoverlapping patch selection reduces computational complexity, without compromising image quality. Experimental results show significant improvement in terms of subjective and objective quality than other popular learning or interpolation based method.Our method showsrobustness on noisy images also.

中文翻译：

---

具有改进边缘信息的基于学习的单图像超分辨率

摘要

本文提出了一种新的基于学习的单图像超分辨率技术，该技术可以将低分辨率（LR）图像单次放大到所需的高分辨率（HR）图像。在放大过程中，从外部数据集学习线性映射功能。映射功能将LR补丁转换为其对应的补丁。在大多数基于补丁的学习技术中，重叠区域的平滑度是通过重叠区域的平均值来执行的。结果，反映在相邻LR补丁中的边缘信息并不总是透明地反映在HR补丁中。因此，在我们的技术中，我们在相邻面片中应用了边缘定向的平滑度。沿方向存在边缘，其中二阶导数较低。为此，我们选择了非重叠补丁，并在获得HR补丁后，我们执行了相邻面片的边缘定向平滑度。这样可以使相邻补丁的平滑度更加详细。除了这种不重叠的补丁选择之外，还可以降低计算复杂度，而不会影响图像质量。实验结果表明，与其他流行的基于学习或插值的方法相比，该方法在主观和客观质量方面均取得了显着改善。