A Fast Outlier Image Detection Method Based on Thumbnails of Space borne SAR Images

Due to the SAR imaging mechanism and other factors such as satellite-ground abnormality, images generated from satellite data occasionally encounter quality problems, such as defocusing, ghosting, interference, and dark or light spots. If the outlier images caused by these qualityproblems are not detected and filtered in time, the subsequent processing may be ineffective, causing resource waste and decreasing the entire efficiency. The data amount of SAR image products is so huge that the computational burden is high. Detecting outlier images on the thumbnail scale will greatly lower the processing time and raise the comprehensive efficiency of the entire application chain. To detect outlier images, the deep-learning-based method is the optimal choice because of the self-learn and high-automation ability. Therefore, aiming to detect outlier images using thumbnails of space borne SAR images, this paper selects the Darknet model as the backbone and employs two architectures of different parameter scales, Darknet53 and Darknet19,to conduct the experiments. Experimental results validate the feasibility of outlier image detection only using thumbnails, and the detection accuracy basically meets the requirement when the efficiency is much higher than the current manual outlier image detection scheme.