Research And Implementation Of Core Images Denoising Algorithm Based On Generative Adversarial Network

Core cast sheet images(hereafter referred to as core images)are highly fidelity images of the physical core,which can reflect the original appearance of the subsurface and play an important role in sedimentary environment inference,reservoir physical property prediction,and oil and gas analysis.The core images can be processed to better understand the geological structure of the core.However,core images are easily contaminated by noise when they are acquired,and image denoising becomes a primary problem.At present,image denoising algorithms based on deep learning can solve the problems of large computation,blurred edges,loss of details,and poor quality of generated images by traditional algorithms,and thus are widely used.In order to overcome the current challenges of core image denoising,the research and implementation of core image denoising algorithms based on generative adversarial networks have two main works as follows:(1)A global residual-based core image denoising algorithm is investigated and implemented.The algorithm is optimized in two aspects: the full pre-activation residual module and a global jump connection from input to output are used in the generator network to optimize the training process and solve the problem of gradient disappearance;the gradient penalty mechanism is introduced in the adversarial loss function,and the adversarial loss,pixel loss,perceptual loss and smoothing loss are weighted and fused as the loss function.An optimized composite perceptual loss function is obtained to improve the model denoising ability from pixel,feature and semantic levels.The experimental analysis shows that the network can effectively remove noise in both J and S oil fields,and the generated images retain the integrity of structural features in core images,obtaining better visual effects and effectively verifying the feasibility of the improved method.(2)A core image denoising algorithm based on dual frequency features and attention residuals is investigated and implemented.The algorithm performs the following operations to further recover texture details of core images and enhance the ability of generative adversarial networks to distinguish noisy features:(1)improving the feature extraction capability by adding octave convolution in the generator,which divides the feature map into high-frequency and low-frequency parts;(2)The channel attention module is embedded at the end of the residual structure to improve the sensitivity of the network to noise and avoid the loss of details during feature extraction.(3)The discriminator uses Patch GAN to guide the generator to learn core images with better image quality.The experimental results show that the network can effectively remove noise in both J and S oil fields,and the generated denoised images can obtain excellent image quality,recover more texture details,and perform well in both visual effects and evaluation indexes,effectively verifying the advantages of the improved method.The comprehensive experimental results show that,for the core image denoising problem,the proposed algorithm in this thesis has the following advantages compared with CBM3 D,WNNM,GCGAN and FFDNet,etc.: on the one hand,it improves the training speed while efficiently eliminating noise and obtaining more realistic core images;on the other hand,it reduces the cost of artificial core image denoising and provides important technical support for subsequent core image processing,which has certain theoretical significance and application value,and also has important practical significance for the subsequent exploration and development in the field of oil and gas fields.