PARAMETER OPTIMIZATION OF RECONSTRUCTION AND REGULARIZATION MULTI FRAME SUPER RESOLUTION SCHEME

• Main Author: Zahi Ulul Azmi, Ahmad
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/51997
• Institution: Institut Teknologi Bandung

Multi Frame Super Resolution is a method for increasing the detail of the sub-pixels in an image or to obtain information at a higher spatial frequency. This method uses several images with aliasing variations or differences in the sub-pixel section to form an image with a higher spatial resolution. There are two methods to be compared, namely reconstruction and regularization. The reconstruction method uses an image registration process, interpolation and restoration. Meanwhile in the regularization method, the inversion of the image acquisition model is used to obtain high resolution images. One of the super resolution implementations is remote sensing imaging on the PROBA-V satellite, the satellite uses imaging with Near Infrared (NIR) channels. Physically, an NIR image sensor has a smaller resolution than a visible light image sensor, therefore it is necessary to increase the resolution of the NIR image to obtain more information in higher spatial frequency. The enhancement with the Super Resolution method can be done with 9 low resolution (LR) images and one high resolution (HR) image as ground truth for evaluating the results of Super Resolution. The HR image has a resolution three times greater than the LR image, where the LR images has a resolution of 128x128 pixels while the HR image is 384x384 pixels. In the nine LR images, there are differences in the histogram, Point Spread Function (PSF), Probability Density Function (PDF), noise, and movement of objects in the image, therefore synthetic images are generated from high resolution images with translation variations and homogeneous PSF and PDF values. PROBA-V data image and synthesis image will be processed using the two super resolution methods developed using Matlab, the output of the algorithm is an image with a resolution of 384x384 pixels. However, in the reconstruction and regularization methods, there are parameters that cannot be determined theoretically, so that genetic algorithm optimization with a specific objective functions as image quality parameter which are Peak Signal-to-Noise Ratio (PSNR) and Structure Similarity Measurement Index (SSIM). The optimized parameter for reconstruction method is size and standard deviation of PSF for deblurring, deblurring iteration, and median filter windowing size. Meanwhile, regularization optimized parameters are regularization weight, BTV window size, alpha factor, and median filter window size. The results of reconstruction method using PROBA-V data input have PSNR and SSIM values of 36.53 and 0.943, meanwhile for the synthetic images the scores are 39.89 and 0.939. Compared to the reconstruction method, for the regularization method with PROBA-V data images, the PSNR and SSIM values are 36.20 and 0.937, while for synthetic images were 0.953 and 41.18. In addition, both Super Resolution methods are evaluated in thermal line images. The images are thermal line objects with width variation on a PCB captured using FLIR E6 camera with an acquisition distance of 33 cm and 55 cm. The image is tested by comparing the result of super resolution Multi Frame and simple bicubic interpolation. Super Resolution result images shown there are increments in image quality which is indicated by lines with 1 mm intervals that clearly separated. This applies to both regularization and reconstruction methods. Meanwhile, in bicubic interpolation result images, the lines with 1 mm intervals are not shown completely separated. The good result of the super resolution can also be seen on the 2mm gradient on the edge of the line. The gradient looks very sharp in the image generated by super resolution based on regularization, although it is slightly blurry in the reconstruction method. However, in bicubic interpolation, the gradient is almost invisible. So, it can be concluded that both super Multi Frame resolution image are able to produce more information about objects in the image or detail at higher spatial frequencies.