| Image resolution is a performance parameter used to evaluate the richness of detail information contained in an image,which represents the ability of an imaging system to truly reflect the detailed information of an object.Compared with low-resolution images,highresolution images generally contain larger pixel density,more texture details and higher reliability,and have higher application value in satellite remote sensing,video surveillance,medical imaging and other fields.As an effective means to improve image resolution,image super-resolution reconstruction technology has the advantages of high efficiency,low cost and wide application,and has been a hot research topic in the field of image processing and computer vision tasks.The thesis mainly studies the application of residual network structure,attention mechanism and generative adversarial network in the field of image super-resolution,and proposes two image super-resolution reconstruction algorithms,which can solve the problem of gradient disappearance and model training instability,and improve the quality of reconstructed images.In order to solve the problem that traditional neural network only use the spatial domain information in the process of image super-resolution reconstruction,resulting in the loss of important details of the generated images.Based on the research of existing image superresolution algorithms in the area of deep learning,the thesis proposes a super-resolution reconstruction algorithm based on the depth residual attention mechanism.Firstly,a deep residual network is constructed by combining local residual learning and global residual learning.The network proposed in the thesis can not only fully excavate the internal features of the image,but also reduce the degradation of the network.Aiming at the residual network,an enhanced channel attention mechanism network is added to the local residual network,which strengthens the ability of neural network to extract high-frequency feature information,reduces the phenomenon of gradient disappearance,and improves the efficiency of information transmission.Aiming at the reconstruction module,different scales of the convolution kernel are used for feature extraction to obtain the image feature information of different scales,and the high-resolution image with rich texture details can be reconstructed.Based on the research of existing image super-resolution algorithms in the area of generative adversarial network,the thesis improves the generator network and the discriminator network respectively.In the generator network,the residual dense block and the improved channel attention mechanism are added to make the image generated by the generator network more similar to the original image,and make the neural network pay more attention to the highfrequency information in the image.In the discriminator network,the thesis uses the relativistic discriminator network instead of the standard discriminator network,so that the real image output by the neural network has a more real probability than the fake image.The activation function in the neural network is analyzed and improved,so that the texture details learned by the network are more detailed.In order to verify the effectiveness of the algorithm,two methods proposed in the thesis are compared with some classical image super-resolution algorithms.The experimental results show that the image reconstructed by the method in the thesis is clearer,the detail texture is more real,and the reconstruction effect can obtain better objective evaluation indexes and subjective visual effects compared with the contrast algorithms. |
