Super-resolution Reconstruction Of Dynamic Scene Images Based On Array Camera

In the field of image processing,super-resolution reconstruction of images is one of the most important research topics.With the development of 3D reconstruction,super-resolution reconstruction,distance measurement and high-speed camera technology,array cameras have become an important research focus in image processing.Microarray cameras have the characteristics of small size,multi views,and the possibility of applying to portable devices.In this paper,a microarray camera is used to capture array images of a dynamic scene.And the super-resolution reconstruction method suitable for dynamic scene array image is further studied.The main work and innovation in this study include:(1)The importance of image super-resolution reconstruction is introduced,and the specific reasons for the formation of low resolution image are analyzed.The basic reconstruction model of array images and single image is briefly introduced.(2)Preprocessing of dynamic scene array images.After analyzing the characteristics of the microarray camera in detail,the distortion correction,registration and sharpening of the dynamic scene array images are carried out.Finally,the dynamic scene images are interpolated to obtain the initial high resolution image.(3)Super-resolution reconstruction of convolution neural network.An improved convolution neural network is proposed in this paper.First,the convolution layer of our convolution neural network is 3 x 3 or 1 x 1 layer,of which 1 x 1 layer is used to improve the network performance.Secondly,the bottleneck structure is applied to reduce the parameter numbers of the nonlinear mapping and improve the nonlinear capability of the whole network.Finally,we use a 3 x 3 deconvolution layer to significantly reduce the number of parameters compared to the deconvolution layer of FSRCNN-s.After the reconstruction of the network,the quality of the image is improved.The experiments prove that our method can effectively use the information between the array images to improve the texture quality of the target image.The convolution neural network proposed in this paper is better than SRCNN and FSRCNN-s.The convolution neural network can further improve the quality of the initial high resolution image,and finally get a better dynamic scene array image.