Dynamic imaging inversion with double deep learning networks for cameras

To achieve high-quality imaging in low-light conditions, a remote sensing camera usually adopts a dynamic time-delay-integration imaging approach, which requires an accurate matching relationship between the optical image field motion and the photo-induced charge transfer. High-frequency motion aberrations still exist due to the measurement frequency limitation of physical attitude and position measurement sensors. However, conventional inversion imaging methods, such as blind deconvolution, can only measure and remove low-frequency motion aberrations. Here, an efficient dynamic inversion imaging algorithm based on double deep learning networks is proposed, which is able to measure and remove high-frequency motion aberrations. To measure high-frequency motion aberrations, we constructed two supervised online deep learning networks, a high-frequency motion aberration inversion learning network (HMAILN) and an optical flow inversion learning network (OFILN). The OFILN can measure the accurate optical flow information, which forms the input training set of the HMAILN. The HMAILN completes the measurement of high-frequency motion aberrations. Finally, the measured high-frequency motion aberrations from the HMAILN were used to construct the motion point spread function for imaging compensation to remove high-frequency motion aberrations. The proposed method was experimentally confirmed, opening the door for the successful implementation of dynamic high-resolution imaging without high-frequency motion aberrations.