Large-Scale Benchmark for Uncooled Infrared Image Deblurring

Abstract

Infrared images are increasingly adopted in various applications. Therefore, motion deblurring for infrared images is also receiving growing interest. However, deep-learning-based deblurring techniques for infrared images have yet to be deeply studied, since there is no publicly available dataset for training and evaluating the networks. In this article, we introduce a large-scale dynamic scene deblurring dataset for microbolometer-based uncooled infrared detectors named uncooled infrared image deblurring (UIRD), which reflects their unique blur characteristics. The dataset is synthetically generated from cooled midwave infrared (MWIR) camera images using a combination of frame interpolation, IR band conversion, and a unique blur accumulation model for the microbolometer. The dataset consists of more than 30k blur-sharp image pairs, and we show the effectiveness of our dataset by showing deblurring results on real uncooled infrared images with the deblurring algorithms trained with our dataset. Our dataset is publicly released to facilitate future research in this area.