부산대학교 도서관

More and more kinds of sensors are used including cameras in the vehicle to proactively address safety issues, either directly or indirectly. Camera failures, such as abnormal frames caused by muzzy, obstruction, and flutter, can lead to system exceptions and even traffic accidents because of their important role in the vehicle's system. We hope to reduce those exception faults by recovering abnormal frames. Therefore, in this paper, we first collect the video from the front-facing camera and define the abnormal frames. Then, this dataset is learned by a cycle generative adversarial network (CycleGAN) to generate more abnormal frames because sufficient samples are needed for better training. Moreover, CycleGAN can also restore the abnormal frames to normal frames, which reduces the system faults. This method can mitigate the consequence of camera failures and also works as a generator of corresponding failure frames.