| Abstract |
Recently, the demand for the enhancement of low-quality images has grown tremendously. The aim of image enhancement techniques is to improve the visual appearance and simultaneously extract valuable information details that can be used for automated image processing applications, such as segmentation and detection. Many types of images, such as astronomical, surveillance, medical, or even real life images that were captured through a camera, or a cell-phone, suffer from low-resolution, degradation effects, poor illuminated regions, and from the existence of severe noise. As a result, enhancing the quality of low-light, low-quality images, is a critical processing step due to its value from both an aesthetics and an information extraction perspective.
In this thesis, we focus on two specific applications of image enhancement techniques, namely, the Super-Resolution and Deconvolution problem (SR-DC), and the Low Light Image Enhancement problem. In both approaches, we utilize the novel framework of the sparse representations and learned appropriate dictionaries, as a prior knowledge that can efficiently represent the enhanced version of a given degraded scene. In our model, the sparse representation of the low-quality image patches, in an appropriate dictionary is used for the approximation of the high-quality images.
Considering the Super-Resolution and Deconvolution problem, we generate two dictionaries, one for the high resolution image parts (high resolution dictionary), and one for the blurry and low-resolution image parts (blurry kernel dictionary). The high resolution dictionary, is trained through a collection of high resolution patches, using a machine learning approach. In addition, the blurry kernel dictionary is created by Gaussian point spread functions (psf's) with different variances. The combination of the two dictionaries creates a single joint dictionary that can efficiently address the super-resolution and semi-blind deconvolution problem. In order to evaluate the performance of the proposed deconvolution and super-resolution method, we performed a comparison in the terms of Peak Signal to Noise Ratio (PSNR), against the standard Bicubic interpolation approach.
We also considered an extended model for the enhancement of low-illumination scenes. We utilize the sparse representation of low light image patches in an appropriate dictionary to approximate the corresponding day-time images. We consider two dictionaries; a night dictionary for low light conditions and a day dictionary for well illuminated conditions. To approximate the generation of low and high illumination image pairs, we generated the day dictionary from patches taken from well exposed images, while the night dictionary is created by extracting appropriate features from under exposed image patches. Lastly, we introduced, a novel scheme in the joint dictionary training phase, using the probabilistic framework of the Restricted Boltzman Machines networks.
Experimental results suggest that the proposed scheme is able to accurately estimate a well illuminated image given a low-illumination version. The effectiveness of our system is evaluated by comparisons against ground truth images, in terms of the Structural Similarity Index (SSIM). When compared to other methods for image night context enhancement, our system achieves better results both quantitatively as well as qualitatively.
|
Search
