[1N3-GS-10] Vision, speech: GAN

[1N3-GS-10-05] Deep image restoration for object detection in compressed dashcam videos

〇Masakazu Tobeta Tobeta¹, Sawa Takamuku¹, Naotake Natori¹, Yoshihiro Honda², Naohiro Hiraiwa², Takashi Mizuno² (1. AISIN SEIKI Co., Ltd, 2. AISIN AW Co., Ltd)

Keywords:automobile, computer vision, deep learning, image restoration, object detection

Given remarkable developments of information and communications technology (ICT) such as cloud computing and 5G networks, quality of service (QoS) in the area of computer vision for automobiles is expected to improve by shifting from standalone computing to cloud computing.However, as dashcam videos transmitted to clouds via 5G networks still have to be compressed with a relatively high compression ratio because of the latency / cost, the lack of image quality for taking advantage of high efficiency of clouds is a great deal of concern.In this study, we apply a deep image restoration method based on adversarial learning for object detection in compressed videos and evaluate its effectiveness in an assumed cloud environment.