Research On ESRGAN Improvement And Its Super-resolution Application To Rock Micro-image

Rock microscopic images can reflect the distribution of oil and gas reservoirs and have high application value for oil exploration and other industries,but due to the limitation of equipment,environment and other factors,rock microscopic images usually have low resolution and loss of detail features.With the development of deep learning and neural networks,the learning ability of various deep learning network models for data has been enhanced,and their experimental results in the corresponding fields are usually better than traditional methods.In the field of machine vision and image processing,image super-resolution is commonly used as a method to improve image resolution and detail features,and a neural network model trained by image data can achieve high-resolution images with rich detail features recovered from low-resolution images.Therefore,in this paper,we choose a deep learning approach to improve the network model based on the enhanced generative adversarial network’s image super-resolution reconstruction algorithm(ESRGAN),and apply the improved algorithm to the super-resolution reconstruction of rock microscopic images,with two main works as follows.(1)An improved method of ESRGAN generation network based on multi-scale feature fusion and an improved method of discriminator based on U-net are proposed.The U-Net discriminator can discriminate the authenticity of individual pixel values at the pixel level,which can ensure the accuracy of the overall authenticity of the image and focus on its detailed features.After experiments on rock microscopic images,the PSNR index and SSIM index of the images are effectively improved.(2)A super-resolution processing network performance optimization method based on the characteristics of rock microscopic images is proposed.In conventional super-resolution studies,the objects are basically the commonly used classical data sets,such as DIV2 K,BSD100,Set5,Set12,etc.,which often consist of images with a wide range of features.In contrast,only detailed features such as particles and pores exist in rock microscopic images.In order to enhance the effect of ESRGAN in reconstructing rock pore features,porosity is introduced as a new constraint term.This constraint term consists of the original high-resolution image porosity and the generated image porosity.In this study,the porosity constraint term is introduced to realize the improvement and optimization of the loss function,and the experimental results show that the network-generated high-resolution images have better performance in terms of evaluation index and visual effect after the introduction of the porosity constraint term,and their porosity measurement results are also significantly improved.In summary,this paper proposes two improvement methods to enhance the evaluation indexes and visual effects of the network reconstructed images by improving and optimizing the network structure of ESRGAN.The theoretical analysis lays down the feasibility of the methods,and the experimental results confirm the effectiveness of the methods.