Research On Dual-Focal Camera Continuous Digital Algorithm

In recent years,dual-focal imaging has played a pivotal role in the field of aerospace remote sensing and smart phone applications.Dual-focal imaging technology(especially dual-resolution cameras)occupies less volume and weight than traditional optical zoom cameras,and combined with image processing technology can have a certain zoom capability.Because dual-focal imaging can only capture images with fixed two zoom magnifications,a digital zoom software algorithm needs to be designed.When the user gives any zoom magnification,the field of view angle and imaging magnification are equivalent to the optical zoom.The image quality of the digital zoom image is required to be as good as possible;for the texture in the short focus field of view but outside the telephoto field of view,the algorithm is required to make full use of the texture information in the telephoto field of view to enhance the texture in order to obtain more Good subjective effect and texture sharpness;the algorithm needs to be robust to the parallax of the dual resolution camera.This paper first designs a continuous digital zoom algorithm scheme.This solution is based on the method of feature layer extraction and feature patch matching of convolutional neural network,making full use of the shooting information of two different focal length lenses,transferring the high-resolution information of the telephoto lens image to the repairable area of the short focus image,and repairing the short-focus image afterwards,the digital zoom is gotten.On this basis,this algorithm has been improved,and the main improvements are:The first is to improve the patch matching algorithm on the feature layer,making full use of the advantages of GPU parallel computing,and completing the texture migration through one forward propagation,avoiding the large time overhead caused by the iterative optimization method;the second is to use a trainable threshold network instead of the single threshold processing of the above algorithm,while achieving better threshold robustness,the network also has the ability of single-frame super-resolution and other image enhancements,and it also enhances the quality of unrepairable textures in short focus images.A reasonable real shot dataset was also proposed,which enables the algorithm to have very good zoom processing capabilities for real shot images.Simulation and real experiments show that compared with the traditional interpolation-based zooming scheme and the single-image super-resolution(SR)zooming scheme,the proposed scheme has higher subjective resolution and visual clarity,and higher objective evaluation indicators;when the user gives a zoom factor between the long and short focus lens multiples,the proposed solution can significantly improve the quality of the zoomed image;for textures outside the field of view of the telephoto camera but in the field of view of the short focus camera,proposed effect is better than that of the existing methods;the processing result of proposed method is very robust to the binocular disparity of telephoto and short focus images,and is significantly better than the continuous zoom algorithm based on registration fusion.