Digital Zooming Algorithm For Dual-Camera System Via High-frequency Inpainting

An asymmetrical dual camera systems refers to system constituted by two fixed focal-length camera system.They are becoming more and more popular among many commercial products because they are very economical and robust.In this application,two fixed-lens cameras are expected to provide digital zooming images.Existing related algorithms such as super-solution and image fusion can efficiently use information from both tele-view camera and wide-view camera.But because of the sampling rate inconsistency,especially when the focal lengths of cameras vary largely,the fused images are usually not visually satisfying.A new photo-realistic continuous digital zooming algorithm proposed in this paper is consists of a super-resolution framework and a band-inpainting algorithm.A super-resolution framework has been proposed in this paper.This framework only undergoes a few super-resolution processes to reconstruct the synthetic digital zooming image,which largely reduces the demand on computational capability and memory storage.The proposed algorithm also consists of a band-inpainting algorithm to reconstruct texture areas.The proposed band-inpainting algorithm can overcome the problem that super-resolution algorithm has a poor performance to reconstruct texture areas.Inspired by the iterative back-projection algorithm,a computational efficient image super-resolution model is also proposed on the basis of the image degradation model.It can produce super-resolved images with negligible algorithmic complexity.The result illustrates that proposed digital zooming algorithm combines the advantages of both super-resolution algorithm and image-inpainting algorithm,and can produce state-of-the-art photo-realistic digital zooming images,and 2dB higher than the traditional bicubic interpolation.Although reconstruction accuracy of proposed single-frame image super-rsolution algorithm via convolutional recurrent network is slightly inferior to large,complicated models by about 0.2dB,it also has negligible algorithmic complexity,and show strong robustness in different application scenarios.