Analysis of Remote Sensing Image Super Resolution using Fluid Lenses

Abstract

Käesolevas doktoritöös uuriti nii riist- kui ka tarkvaralisi lahendusi piltide töötlemiseks. Riist¬varalise poole pealt pakuti lahenduseks uudset vedelläätse, milles on dielekt¬rilisest elastomeerist kihilise täituriga membraan otse optilisel teljel. Doktoritöö käigus arendati välja kaks prototüüpi kahe erineva dielektrilisest elastomeerist ki¬hilise täituriga, mille aktiivne ala oli ühel juhul 40 ja teisel 20 mm. Läätse töö vas¬tas elastomeeri deformatsiooni mehaanikale ja suhtelistele muutustele fookuskau¬guses. Muutuste demonstreerimiseks meniskis ja läätse fookuskauguse mõõtmiseks kasutati laserkiirt. Katseandmetest selgub, et muutuste tekitamiseks on vajalik pinge vahemikus 50 kuni 750 volti. Tarkvaralise poole pealt pakuti uut satelliitpiltide parandamise süsteemi. Paku¬tud süsteem jagas mürase sisendpildi DT-CWT laineteisenduse abil mitmeteks sagedusalamribadeks. Pärast müra eemaldamist LA-BSF funktsiooni abil suu¬rendati pildi resolutsiooni DWT-ga ja kõrgsagedusliku alamriba piltide interpo¬leerimisega. Interpoleerimise faktor algsele pildile oli pool sellest, mida kasutati kõrgsagedusliku alamriba piltide interpoleerimisel ning superresolutsiooniga pilt rekonst¬rueeriti IDWT abil. Käesolevas doktoritöös pakuti tarkvaraliseks lahenduseks uudset sõnastiku baasil töötavat super-resolutsiooni (SR) meetodit, milles luuakse paarid suure resolutsiooniga (HR) ja madala resolut-siooniga (LR) piltidest. Kõigepealt jagati vastava sõnastiku loomiseks HR ja LR paarid omakorda osadeks. Esialgse HR kujutise saamiseks LR sisendpildist kombineeriti HR osi. HR osad valiti sõnastikust nii, et neile vastavad LR osad oleksid võimalikult lähedased sisendiks olevale LR pil¬dile. Iga valitud HR osa heledust korrigeeriti, et vähendada kõrvuti asuvate osade heleduse erine¬vusi superresolutsiooniga pildil. Plokkide efekti vähendamiseks ar¬vutati saadud SR pildi keskmine ning bikuupinterpolatsiooni pilt. Lisaks pakuti käesolevas doktoritöös välja kernelid, mille tulemusel on võimalik saadud SR pilte teravamaks muuta. Pakutud kernelite tõhususe tõestamiseks kasutati [83] ja [50] poolt pakutud resolutsiooni parandamise meetodeid. Superreso¬lutsiooniga pilt saadi iga kerneli tehtud HR pildi kombineerimise teel alpha blen¬dingu meetodit kasutades. Pakutud meetodeid ja kerneleid võrreldi erinevate tavaliste ja kaasaegsete meetoditega. Kvantita-tiivsetest katseandmetest ja saadud piltide kvaliteedi visuaal¬sest hindamisest selgus, et pakutud meetodid on tavaliste kaasaegsete meetoditega võrreldes paremad.

In this thesis, a study of both hardware and software solutions for image enhance¬ment has been done. On the hardware side, a new liquid lens design with a DESA membrane located directly in the optical path has been demonstrated. Two pro¬totypes with two different DESA, which have a 40 and 20 mm active area in diameter, were developed. The lens performance was consistent with the mechan¬ics of elastomer deformation and relative focal length changes. A laser beam was used to show the change in the meniscus and to measure the focal length of the lens. The experimental results demonstrate that voltage in the range of 50 to 750 V is required to create change in the meniscus. On the software side, a new satellite image enhancement system was proposed. The proposed technique decomposed the noisy input image into various frequency subbands by using DT-CWT. After removing the noise by applying the LA-BSF technique, its resolution was enhanced by employing DWT and interpolating the high-frequency subband images. An original image was interpolated with half of the interpolation factor used for interpolating the high-frequency subband images, and the super-resolved image was reconstructed by using IDWT. A novel single-image SR method based on a generating dictionary from pairs of HR and their corresponding LR images was proposed. Firstly, HR and LR pairs were divided into patches in order to make HR and LR dictionaries respectively. The initial HR representation of an input LR image was calculated by combining the HR patches. These HR patches are chosen from the HR dictionary corre-sponding to the LR patches that have the closest distance to the patches of the in¬put LR image. Each selected HR patch was processed further by passing through an illumination enhancement processing order to reduce the noticeable change of illumination between neighbor patches in the super-resolved image. In order to reduce the blocking effect, the average of the obtained SR image and the bicubic interpolated image was calculated. The new kernels for sampling have also been proposed. The kernels can improve the SR by resulting in a sharper image. In order to demonstrate the effectiveness of the proposed kernels, the techniques from [83] and [50] for resolution enhance¬ment were adopted. The super-resolved image was achieved by combining the HR images produced by each of the proposed kernels using the alpha blending tech-nique. The proposed techniques and kernels are compared with various conventional and state-of-the-art techniques, and the quantitative test results and visual results on the final image quality show the superiority of the proposed techniques and ker¬nels over conventional and state-of-art techniques