【摘要】 针对传统的超分辨率重建算法在处理光照不均图像时会出现图像失真、边缘模糊等问题,在原有的极深超分辨率(VDSR)重建方法基础上提出一种光照不均图像的超分辨率重建方法。首先,采用自门控Swish激活函数代替常用的ReLU激活函数,解决了随着网络层数加深出现的过拟合问题,可以更好地学习映射关系;然后,在网络结构中提出一种简洁紧凑型的局部残差网络,在保证网络层数的同时能学习更多的图像细节信息,很好地解决了VDSR中由于图像多次传输出现的信息丢失问题;最后,在网络末端使用反卷积获得高分辨率图像。通过实验证明该方法对光照不均图像重建可以获得更高的峰值信噪比(PSNR)和结构相似度(SSIM)。更多还原

【Abstract】 In order to get rid of the image distortion and edge blur occurred when processing the images with uneven illumination by the traditional super-resolution reconstruction algorithm,a super-resolution reconstruction method for images with uneven illumination is proposed based on a very deep super-resolution(VDSR) reconstruction method. A self-gated Swish activation function is adopted instead of the commonly used ReLU activation function,which gets rid of the over-fitting generated as the deepening of network layer and facilitates better learning of mapping relationship. And then,a simple and compact local residual network is adopted in the network structure. It can be used to learn much image detail information while ensuring enough network layers,which can properly solve information loss caused by repeated image transmission in VDSR. A high-resolution image is obtained by deconvolution used at the end-point of the network. It is verified by experiments that this method can achieve higher PSNR(peak signal-to-noise ratio)and SSIM(structural similarity)in the reconstruction of images with uneven illumination.更多还原