Super Change with Super Resolution Algorithms
December 2, 2011
School of Engineering, Bengaluru
Many of us use CMOS (Complementary Metal Oxide Semiconductors) devices. We may not be aware, but these semiconductors are embedded in our cell phones, digital cameras and digital video cameras. Even in some vehicles and toys.
Security cameras, factory inspection systems, medical equipment also make use of CMOS for image processing.
These semiconductors are somewhat expensive; however a large number is needed in a device to ensure high-quality picture resolution. Where price is an issue, the technique of Image Super Resolution comes to the rescue.
This software technique offers a viable alternative, enhancing the resolution of an imaging system through digital signal processing. A popular area of research today in universities and in the semiconductor industry, the emphasis is on creating effective super-resolution algorithms.
The Amrita School of Engineering at Bengaluru is at the forefront of some of this research. S. Ravishankar, Professor, Department of Electronics and Communication Engineering, is focused on producing an error-minimized output.
He described the reasons for developing a new algorithm for Image Super Resolution.
“The image enhancement feature integrated in a CMOS system utilizes sophisticated software algorithms known as ‘Image Super Resolution Algorithms.’ These provide high resolution images. In the absence of these algorithms, we would require a large number of CMOS sensors, and these are difficult to fabricate within the miniaturization constraint.”
“Therefore, many imaging systems are designed with a less dense detector array which causes aliasing during image acquisition. In addition, there is a loss of high frequency details due to low-resolution point spread function and optical blurring due to motion. In such cases, one has to resort to suitable image processing techniques to enhance the resolution.”
Ravishankar presented a paper titled Image Super-Resolution using Sparse Image and Singular Values as Prior* at the 14th International Conference on Computer Analysis of Images and Patterns at University of Seville, Spain during August 29-31, 2011.
A paper titled Single Image Super Resolution Using Sparse Image and GLCM Statistics as Priors** was also presented at the International Conference on Signal Processing and Image Engineering, during July 6-8, 2011, at the Imperial College in London.
“Given a low-resolution image, and training patches of high-resolution and low-resolution patches obtained from sampling similar images, we reconstruct the high-resolution version of a low-resolution image,” he elaborated.
Given his efforts for rapid and relevant changes in the area of Image Super Resolution, we know we can definitely look forward to a feasible, error-minimized outcome.
* Paper Abstract – Sparse representation of images finds many applications in image processing and computer vision. Recently various attempts have been made to regulararize the ill posed inverse problem of motion free image super resolution, using sparse representation of low resolution image patches. A few of these approaches use only sparse representation of image patches for super resolving low resolution images. However the proposed method is different from these approaches reported earlier in the literature according to the method of dictionary training, feature extraction, from training database images and regularization techniques. In this work we have used singular values for regularization of the cost function of the image super resolution model. Matching optimal direction algorithm (MOD) has been used in the proposed algorithm for obtaining HR and LR dictionaries from training image patches. Using these two dictionaries the given low resolution input image is super resolved. The results of the proposed algorithm showed visual, PSNR, RMSE and SSIM improvement over other methods.
** Paper Abstract – Sparse representation of images ﬁnds many applications in image processing and computer vision. Recently various attempts have been made to regularize the ill-posed
inverse problem of motion free image super resolution using sparse representation of low resolution image patches. However the proposed method in this paper is different from the previous approaches reported in the literature in terms of method of dictionary training and feature extraction from the trained data
base images. Gray Level Co-Occurrence Matrix(GLCM) is a proven method for extracting image statistical features, which are used mainly for image classiﬁcation, segmentation etc. In the present work we have extracted GLCM parameters for regularization of the data ﬁtting term of the cost function of the image super resolution model. We used Matching Optimal Directions (MOD) algorithm for obtaining high resolution and low resolution dictionaries from training image patches and seek the sparse representation of low resolution input image patch using low resolution dictionary and then obtain high resolution image patch from high resolution dictionary. The results of the proposed algorithm showed visual, PSNR, RMSE,
and SSIM improvements over other super resolution methods.