| Image super-resolution reconstruction is currently a hot and challenging research direction.With the development of artificial intelligence and deep learning,various super-resolution reconstruction networks based on convolutional neural networks emerge in an endless stream,and the resolution of the reconstructed image begins to approach the original image,related technologies have been widely used in medical and remote sensing image processing,image quality enhancement and repair,security monitoring and other social and livelihood fields.However,the reconstructed images by these networks have problems such as loss of high-frequency information and lack of texture details,which lead to bottlenecks in technological development.Therefore,some researchers have turned their attention to the generative adversarial network and introduced it into the field of image super-resolution reconstruction,using the generative network and the discriminative network to play against each other to reconstruct images that conform to human visual perception.Although more detail and texture are recovered,but there are still problems such as lack of detail information and unclear picture quality.To further improve the reconstructed image quality,we make architectural improvements on the basis of the most representative super-resolution reconstruction networks based on generative adversarial network:(1)For the generative network,the residual feature aggregation framework is firstly applied to the backbone network to make full use of features at different levels;secondly,a lightweight spatial attention mechanism module is added at the end of each dense block,allowing the network to base on the spatial content the receptive field is adjusted to aggregate more representative features;finally,the LPIPS score is introduced into the perceptual loss to make the loss function more in line with the human visual comparison mechanism.(2)For the discriminant network,this paper proposes a discriminant network composed of symmetric encoder and decoder by borrowing the structure of U-Net network.It uses the encoder module to compare the differences between images,and the decoder module to compare the gaps between pixels,providing feedback to the generative network from two dimensions to improve the authenticity of the reconstructed images.By adjusting the network architecture and loss function of the generative and discriminative networks,the images reconstructed by our network are both realistic and full of details.After comparison of ablation experiments,it proves the effectiveness of the improvement proposed in this paper;after three subjective and objective indicators data display analysis and visual effect comparison of specific details of the image,the performance of the image super-resolution reconstruction network based on improved generative adversarial network proposed in this paper is illustrated in an advanced position. |
