一种注意力机制与SRGAN网络结合的超分辨率算法

Super-resolution Algorithm Combining Attention Mechanism with SRGAN Network

超分辨率是指将一张低分辨率图片转换成高分辨率,在军事领域、工业领域等都有着重要作用.基于生成对抗网络(GAN)的超分变率方法,主要是根据生成对抗网络原理,由生成器生成一张伪高分辨率图片,再由鉴别器计算这张图片与真实高分辨率图片的差值,来衡量这张图片的真实程度.本文基于SRGAN(Super-resolution Generative Adversarial Network)网络主要进行了3点改进:(1)引入了注意力通道机制,即在SRGAN网络中加入CA(Channel Attention)模块,同时增加网络深度以更好的表达高频特征;(2)删除原有的BN(Batch Normalization)层以提升网络性能;(3)修改损失函数,以减少噪声对图片的影响.通过实验表明,本文所采用的方法改善了伪影问题,在Set5、Set10、BSD100测试集上均提升了PSNR(峰值信噪比).

Super resolution is the conversion of a low resolution image into a high resolution image. It plays an important role in military and industrial fields. The super-resolution method based on Generative Adversarial Network( GAN) is mainly based on the principle of Generative Adversarial Network. The generator generates a pseudo high-resolution image,and then the discriminator calculates the difference between this image and the real high-resolution image Value to measure the true degree of this picture. Based on the improvement of the SRGAN network,three major improvements were made in this paper. First: this paper introduces the attention channel mechanism,that is,adding the CA( Channel Attention) module to the SRGAN network,and increasing the network depth to better express high-frequency features;Second: Delete the original BN( Batch Normalization) layer to improve network performance;Third:modify the loss function to reduce the impact of noise on the picture. experiments show that the method used in this article improves the artifact problem and improves the PSNR( peak signal to noise ratio) on the Set5,Set10,BSD100 test et.