Single-Frame Image Super-resolution through Contourlet Learning
- PDF / 1,388,989 Bytes
- 11 Pages / 600.03 x 792 pts Page_size
- 85 Downloads / 174 Views
Single-Frame Image Super-resolution through Contourlet Learning C. V. Jiji and Subhasis Chaudhuri Department of Electrical Engineering, Indian Institute of Technology Bombay, Mumbai 400076, India Received 26 November 2004; Revised 22 March 2005; Accepted 5 April 2005 We propose a learning-based, single-image super-resolution reconstruction technique using the contourlet transform, which is capable of capturing the smoothness along contours making use of directional decompositions. The contourlet coefficients at finer scales of the unknown high-resolution image are learned locally from a set of high-resolution training images, the inverse contourlet transform of which recovers the super-resolved image. In effect, we learn the high-resolution representation of an oriented edge primitive from the training data. Our experiments show that the proposed approach outperforms standard interpolation techniques as well as a standard (Cartesian) wavelet-based learning both visually and in terms of the PSNR values, especially for images with arbitrarily oriented edges. Copyright © 2006 Hindawi Publishing Corporation. All rights reserved.
1.
INTRODUCTION
In most imaging applications, images with high spatial resolution are desired and often required. Resolution enhancement from a single observation using image interpolation techniques is of limited application because of the aliasing present in the low-resolution (LR) image. Super-resolution refers to the process of producing a high spatial resolution image than what is afforded by the physical sensor through postprocessing, making use of one or more lowresolution observations. It includes upsampling the image, thereby increasing the maximum spatial frequency, and removing degradations that arise during the image capture, namely, aliasing and blurring. In general, there are two classes of super-resolution techniques: reconstruction-based and learning-based. In reconstruction-based techniques the high-resolution (HR) image is recovered from several lowresolution observations of the input, but in learning-based super-resolution algorithms a database of several other images are used to obtain the high-resolution image. The single-frame image super-resolution problem arises in several practical situations. In many biometric databases, a large number of images of similar content, shape, and size are available. For example, in investigative criminology one has available face and fingerprint databases. These are often taken in a controlled environment. The question we ask is that if one encounters a poor-quality input image, can it be enhanced using the knowledge of the properties of the database images. Thus, the basic problem that we solve in
this paper is as follows. Given a single low-resolution input image and a database of several high-resolution images, we obtain a high-resolution output. We make use of the recently proposed contourlet transform [1] to learn the best features from the database of high-resolution images while upsampling the input image. The features we learn from t
Data Loading...
