This paper aims to reform a scene-based non-uniformity correction (SBNUC) algorithm to speed up the convergence and improve the non-uniformity correction (NUC) performance in infrared (IR) videos. The proposed algorithm is based on inner-frame updates in addition to the inter-frame updates in the existing in-ter-frame registration-based least mean square (IRLMS) algorithm. Even though the IRLMS had been pro-posed to improve the two representative SBNUC algorithms such LMS and MCA, it also has a limitation due to its slow convergence speed. The slow speed is mainly due to the inter-frame updates in the adopted sto-chastic gradient descent optimization. Accordingly, the IRLMS requires many frames to be converged for the intermittent motion. In this paper, we propose a NUC algorithm by adopting the inner-frame stochastic gradi-ent descent optimization in addition to the inter-frame one so that we can improve the performance of NUC as well as the convergence speed. Furthermore, we divide the residual into the meaningful and meaningless ones and apply the NUC only to the meaningful one, to reduce the NUC errors in the local motion area and near the strong edges. The proposed algorithm can provide the faster and more reliable fixed-pattern noise (FPN) reduction, while maintaining the advantages of IRLMS such as low computational complexity, small storage requirement, and ability to capture temporal drifts of the non-uniformity of IR focal plane array (IRFPA). The experimental results show that the proposed algorithm considerably outperforms the existing IRLMS algorithm.
본 논문에서 우리는 최근의 매우 실용적인 IRLMS 알고리즘의 성능을 개선하기 위해 기존 프레임 간 업데이트 방식에 프레임 내 업데이트를 도입하는 최적화 기법을 제안한다. 프레임 내 업데이트는 IRLMS가 더욱 빨리 FPN을 제거할 수 있도록 해주고, 프레임 간 업데이트 방식 보다 더욱 높은 성능을 가질 수 있도록 해준다. 이러한 프레임 내 업데이트를 안정적으로 수행하기 위해 local motion에 의한 meaningless residual과 sub-pixel motion에 의한 edge 주변의 meaningless residual을 동시에 제거해주는 meaningless residual removal (MRR)를 사용하고, 프레임 내 업데이트 시 residual의 pair matching을 위해 anti-propagation masking으로 overlap되지 않은 영역을 제외시켜준다.
가상 및 실제 데이터에서 제안 알고리듬은 기존 IRLMS의 수렴속도와 성능을 개선하는 결과를 확인할 수 있었고, 단속적인 motion이 존재하는 상황에서도 향상된 결과를 얻을 수 있었다.
