Blind Motion Deblurring for Satellite Image using Convolutional Neural Network

KOMPSAT-3A was launched in March 2015, and has been performing normal operations. However, there is an inaccurate line-rate control in KOMPSAT-3A that causes motion blur. This induces an along-track modulation transfer function (MTF) value that is smaller than the across-track MTF value. The motion blur degrades image quality and tends to deteriorate the performance of various computer vision algorithms such as object detection, segmentation, feature detection, and matching. Thus, in order to enhance the visual quality of blurry images, a motion deblurring algorithm is necessary. However, there are few studies about solving motion blur of satellite images. Most studies have been focused on the motion blur of general images taken by smart phones, action cameras, and DSLRs. To solve the satellite image problem, we propose a model for motion deblurring of satellite images and generate training data using the proposed model and KOMPSAT-3A data. In addition, we trained a residual and dense block-based convolutional neural network (CNN). To verify the residual and dense block-based CNN, we utilized simulated blurred images and actual blurry KOMPSAT-3A images. As a result, we confirmed that the proposed method reliably removes more motion blur than do conventional CNN-based motion deblur methods. Additionally, our method made the along-track and across-track MTF values similar. In addition, overall the MTF values increased. The results show that effective learning is possible for use in reducing motion deblurring of satellite images effectively.