Bifocal camera continuous digital zooming method based on convolutional neural network model

Abstract

The invention discloses a bifocal camera continuous digital zooming method based on a convolutional neural network model. The bifocal camera continuous digital zooming method includes: preprocessing the public data set to obtain high-resolution images, low-resolution images and corresponding reference images with the same size to form image pairs as a training set; establishing a convolutional neural network model to iteratively train the training set for a preset number of times; and inputting a low-resolution image obtained by the short-focus camera and a high-resolution image and zoom magnification obtained by the long-focus camera, and outputting a zoomed image by using a bicubic interpolation method through the convolutional neural network model trained by using the cut up-sampled image and the high-resolution image. According to the method, the convolutional neural network model is utilized to realize digital zoom image synthesis with continuous multiplying power, and compared with an existing continuous digital zoom method, rich texture details provided by a high-resolution image obtained by a telephoto camera can be more effectively utilized.

Description

technical field [0001] The invention belongs to a bifocal camera continuous digital zoom method in the field of digital image processing, and relates to a bifocal camera continuous digital zoom method using a convolutional neural network model. Background technique [0002] Considering the cost, size and reliability of the imaging system, space cameras and smart phones are rarely equipped with optical zoom lenses, but fixed-focus cameras tend to be used. The short-focus camera has a large field of view, but the detail resolution is insufficient, while the telephoto camera can take high-resolution pictures and obtain a lot of detail information, but the field of view is small. Therefore, using two fixed-focus cameras with different focal lengths to form an asymmetric optical system to simulate an optical zoom camera has more and more application scenarios. But such a system introduces a new problem: how to design an efficient algorithm, starting from the input high-resolutio...

AI Technical Summary

Problems solved by technology

But such a system introduces a new problem: how to design an efficient algorithm, starting from the input high-resolution image of the telephoto camera and low-resolution image of the short-focus camera, digitally zooms, and produces continuous images similar to optical zoom visual effects. Digital Zoom Algorithm

[0003] In the process of realizing the continuous digital zoom of bifocal cameras, the following technical difficulties are mainly faced: First, the traditional image fusion algorithm cannot transfer the texture information obtained by the telephoto camera to the outside of the telephoto field of view, resulting in only short-focus Poor texture restoration in areas captured by the camera

Second, the convolutional neural network-based single-frame image super-resolution algorithm that has been deeply studied and widely used can only improve the image quality obtained by short-focus cameras to a certain extent, but cannot effectively use the texture details of long-focus images; The image super-resolution algorithm based on the reference image, the algorithm effect is very dependent on the registration effect of the reference image and the short-focus image. For a fixed value, only fixed magnification zoom can be achieved, and the traditional digital zoom algorithm based on block matching takes too long, so it is difficult to propose a continuous magnification zoom algorithm that can effectively use the texture details obtained by the telephoto camera

Images