SAR 영상 탐지식별 네트워크 벤치마크 및 위상 오차에 의한 열화 분석 (Benchmarking SAR Target Detection Networks and Analysis of Degradation Depending on the Phase Error)

서지정보

· 발행기관 : 한국전자파학회
· 수록지 정보 : 한국전자파학회 논문지 / 35권 / 10호 / 770 ~ 779페이지
· 저자명 : 백민영, 옥재우, 신희섭

초록

SAR 영상은전천후주야간관측에유리하지만광학영상에비해데이터획득및분석이어렵다는단점이있다. 지속적인감시정찰및모니터링의목적으로SAR 영상의활용과더불어딥러닝연구에대한관심이높아지면서모델구축을위한고품질의학습데이터셋에대한필요성역시높아지고있다. 특히SAR 영상에서의탐지식별의경우육안분석이비교적용이하고라벨링의난이도가낮은해상선박에대한데이터셋이대부분이고그외표적에대해서는부족한실정이다. 따라서본연구에서는지상군사표적에대한탐지식별데이터셋을구축하고활용해보고자하였다. 검증된SAR 데이터셋중하나인MSTAR 데이터셋을기반으로새로운탐지데이터셋을구축하였고다양한탐지모델을적용해데이터셋을검증함과동시에지상표적탐지에대한벤치마크를수행한결과전반적으로mAP 0.8 이상의좋은탐지성능을도출하였다. 추가적으로, 구축한탐지모델을실제항공SAR 영상등에서활용할경우를대비해평가영상에위상오차가포함되는경우에대한분석까지수행하였다. 영상및탐지성능열화관점에서탐지성능이크게하락하지않는위상오차에대한허용범위를판단할수있었다. 이를바탕으로평가영상획득시요동제어를위한자료로활용할수있고이외에다양한영상오차를고려하여더욱강건한데이터셋및모델구축에대한기반이될수있을것으로기대한다.

영어초록

SAR images are more difficult to analyze and acquire than optical images; therefore, high-quality datasets for deep learning models are insufficient. In this study, a new detection dataset was generated based on the MSTAR dataset to train detection networks for ground military targets. To verify the dataset and analyze the performance of the ground target detection network in SAR images, various mod els were used to compare the detection ability, resulting in an overall good performance of mAP 0.8 or higher. In addition, we analyzed the performance change trends if phase errors were included in the test images. In terms of image and detection performance degrada tion, we determined an acceptable range for the phase error that did not significantly reduce the detection performance. This is expected to contribute to the creation of more robust networks when building SAR image datasets and models in the future.

참고자료

· 없음