Image Resolution Enhancement Method Based on Feature Space

In this paper, an approach for resolution enhancement of color image based on feature space method is proposed. Many images, such as landscape, natural scene and so on, are fractal, and they can often be assumed as self-correlation. Focusing on this point, we describe the image by feature space. This method reflects the characteristics of color image, and realizes the image interpolation by using the feature space. Experiment result shows that this method can enhance the image resolution more effectively comparing with nearest neighbor interpolation, bilinear interpolation. Introduction Image resolution enhancement is a method of signal processing and image processing. It can transform an existing low resolution image into high resolution image by using the software algorithm. In video monitoring, image printing, forensic analysis, medical image processing, satellite imaging and other fields, image resolution enhancement has been applied widely. Image resolution enhancement related to some basic problem in many fields such as image processing, computer vision, optimization theory. Resolution enhancement has become one of the hot spots in the research field and it has important significance to image processing [1]. Image resolution enhancement is an image data regeneration process within a given range of space. The more image data can be estimated by the finite discrete image data accurately to construct observation images with higher resolution to reflect the real scene. Advanced resolution enhancement technique can effectively enhance the image resolution, so it is widely used in satellite remote sensing, material analysis, medical diagnostic, traffic management, criminal investigation and so on. Resolution enhancement is divided into the traditional interpolation methods and modern interpolation method. The traditional interpolation method is also known as linear interpolation that uses the known image pixel gray value and special interpolation function to calculate the unknown pixel gray value. The typical linear interpolation methods are nearest neighbor interpolation, bilinear interpolation, bicubic interpolation, polynomial interpolation and spline interpolation algorithm [2-6]. The high frequency information of the image is inhibited by traditional interpolation method, so the blur or aliasing phenomenon would appear in the edge region of restoration images, and it is not ideal to restore the edge and texture features of the image. The modern interpolation method is also called nonlinear interpolation, and it introduces the fractal topology, wavelet analysis, partial differential equations, nonlinear optimization theory etc. into digital image processing [7-10]. The modern interpolation methods can use different interpolation method in different regions of image according to the regulation of the image spectrum weighted coefficient. In this paper, the local images are cut out form the color image, such as landscape image, texture image, building image and so on. From these local images, a similar pattern can be extracted by a specific feature space, and the pattern can express the image compactly. Using this pattern we propose a method to enhance the image resolution. 6th International Conference on Electronic, Mechanical, Information and Management (EMIM 2016) © 2016. The authors Published by Atlantis Press 715 Image Pattern Extracted by Feature Space The color images are fractal generally, so the image pattern can be extracted by feature space to express the images. As shown in Fig. 1, sample windows are cut out from image, and then the image vector is generated by scanning the sample window. The image vector can be given as: [ ] [ ] (1) where R is the number of sample windows, m and n are the width and height of sample window, f(x)mn is the gray value in position(m,n), X is the image vector. Then the variance matrix of image vector X can be obtained and using KL transform the Eigen vectors column can be given as: [ ] [ ] (2) where D is the dimension of Feature space. The cumulative proportion is calculated as: ∑ ∑ ⁄ (3) In feature space method, although the dimension is smaller than the rank of the image vector sequence, it can represent the learning samples, because it is often scattered in a particular direction rather than messy. It may also be considered that the learning samples are high correlation. That is, if cumulative proportion is obtained, the eigenvector can be considered as extracted image pattern, and the defective image can restored by extracted image pattern. Resolution Enhancement Using Feature Space Image interpolation is performed by image pattern that expresses the image using eigenvectors. The image vector generated by sample windows is projected onto the Eigen space. The projection point is given as: (4) If there is defect, open hole in sample image, the part of defect is expressed by 0. So,