Proceedings Article | 26 October 2022
KEYWORDS: Turbulence, Deconvolution, Optical flow, Image quality, Detection and tracking algorithms, Image registration, Antennas, Point spread functions, Atmospheric turbulence, Video
As European optronic industries are facing an increasing demand for high-definition imagery, atmospheric turbulence becomes one of the main factors hindering the performance of military equipment used for intelligence, surveillance and reconnaissance. The TURBO project, which is funded by European Defence Agency and has Fraunhofer IOSB, TNO and Adimec as partners, aims to develop a real-time demonstrator capable of selecting the optimal algorithm for the given dataset from a repertoire of software-based turbulence correction algorithms. The work to be presented illustrates Fraunhofer IOSB's contribution to the first stage of the project. It consists of the selection of the most promising candidate methods for the demonstrator, as well as a detailed assessment of their limitations. A total of eight algorithms were selected. The first method performed global registration. For the correction of local motion, optical ow, block matching and iterative image generation methods were applied. To correct blur, we investigated various deconvolution approaches, including Richardson-Lucy approach and a blind deconvolution algorithm, and a Lucky fusion method. Finally, we focused on turbulence correction for scenes with moving content (people, vehicles). All methods were applied to several video sequences, which are recordings from field trials performed under different atmospheric conditions. The resulting corrected sequences were analysed first regarding their quantitative improvement in relation to the uncorrected sequences, and afterwards considering the execution time for each method. Three no-reference metrics were chosen for the quantitative analysis, namely Fisher information, edge variance and total variation. It was demonstrated that the selected methods achieved good correction results in many sequences. In particular optical ow and deconvolution achieved a high increase in image quality. Quantitative analysis was shown to be accurate, in the sense that it agreed systematically with visual perception. It was also demonstrated that certain methods produce good results within a short computation time, deconvolution for instance, which is of great interest for real-time use and it provides a good basis for future work.
