Research On Image Super-Resolution Algorithm Based On Deep Learning

The image resolution is directly reflected in the image definition.Due to the limitations of environment,equipment and other factors in the image acquisition process,the acquired image sometimes lacks high-frequency details,which makes it difficult to obtain enough useful information from the image.The pixels in the image contain a lot of image feature information,so how to extract the useful information in the pixels and establish the relationship between image features is the key research direction to ensure whether the generated image can restore high-definition details.The existing super-resolution reconstruction algorithm is relatively smooth and has serious loss of detail information.In order to improve the quality of the reconstructed image,the existing image super-resolution algorithm based on depth learning is studied,mainly including the following contents:(1)Aiming at the problem that the current image super-resolution reconstruction algorithm is difficult to capture the dependency between feature channels,and the feature propagation in the deep feature extraction network is blocked,an image super-resolution reconstruction network based on residual attention structure is proposed.The network can extract more image features by adding feature channels before activating the function;The residual structure is used to alleviate the problem of network gradient disappearance and explosion.Combine coordinate attention to construct residual attention module,learn the importance of each feature channel,and capture the structure dependency of input feature graph.The 4-fold super-resolution reconstruction results on the Urban100 test set show that the peak signal-to-noise ratio can reach 27.13 d B,effectively improving the utilization of shallow information,and better reconstructing the image details and geometric features,thus improving the quality of super-resolution images.(2)Aiming at the problems of edge blurring and artifacts in super-resolution image reconstruction,a super-resolution algorithm based on generation countermeasure network is proposed.The generator of the algorithm introduces multi-scale residual network,extracts multi-scale features within the residual,and captures the global features of the image by increasing the receptive field of the model.Combining attention mechanism to allocate channel attention resources for features,the dependence between channels and the recovery ability of high-frequency features are enhanced;By removing the batch normalization layer in the residual network,the artifacts caused by the normalization of mean and variance are alleviated.By using the large size convolution kernel in the depth separable convolution replacement discriminator,the computational effort is reduced by 85.32%;The perceptual loss and the absolute average error loss are combined to guide the training process of generating the countermeasure network and avoid the overall smoothness of the reconstructed image caused by feature loss.The simulation results show that the model feature recovery ability of the improved network is significantly improved compared with the mainstream generation countermeasure network algorithm.(3)The algorithm of image super-resolution reconstruction proposed in this paper is applied to the image processing of rock castings,and a prototype system of super-resolution reconstruction of rock castings is designed.The system is developed in a modular way,integrating two kinds of image super-resolution reconstruction algorithms,which can realize the multiresolution reconstruction of rock casting images.The running results of the prototype system show that the prototype system runs stably and the reconstructed image quality is high,which can meet the basic practical needs.The test results show that the two image superresolution reconstruction algorithms proposed in this paper are obviously superior to other algorithms.The two image super-resolution reconstruction algorithms proposed in this paper have a higher degree of image detail recovery and subjective visual effect than existing superresolution reconstruction algorithms in processing image tasks based on bicubic interpolation degradation;The two image reconstruction algorithms are integrated into the designed rock casting image reconstruction prototype system,which performs well in the non degraded rock casting image reconstruction task,indicating that the algorithm proposed in this paper has certain application value.