Decoding motion-blur using variational autoencoder
S. Shah, S. Jha, Saugata Sinha · IET conference proceedings. · 2023
Applications of camera-based systems are increasing day by day due to developments in hardware as well as improvements in software processing. One of the major limiting factors to the extension of these systems in some fields with high-speed applications like automation, self-driving automobiles, etc., is the long exposure time of traditional camera systems to capture images, which results in motion-blurred images. Although this limitation is being tackled by utilizing cameras with higher frame rates (smaller exposure time), such systems increase the development costs of the setup. This paper presents a novel learning-based blur decoding solution, i.e., a combined task of video deblurring and intermediate frame generation from a single motion blur image and its corresponding first sharp image. This method can be used to convert low-frame-rate video to high-frame-rate video by interpolating a single blur frame into multiple sharp frames without modifying the video-acquisition hardware. The proposed network accurately predicts the structure and motion of objects from first clear and blurred frames, respectively, and then uses this information to build a precise sequence of frames. The experimental results show that our algorithm performs substantially well on the GoPro dataset for blur decoding and frame interpolation tasks.