Super-resolution Reconstruction Of Coal Photomicrographs Based On Generative Adversarial Learning

The content of coal macerals is closely related to the physical and chemical properties and industrial uses of coal,and is the basis of coal phase analysis and coal genesis analysis.It is also an important indicator of coal quality and classification.Benefiting from the advantages of high efficiency and strong objectivity,the identification of macerals based on image analysis has gradually become a trend in the industrial application of coal petrology.High-resolution coal photomicrographs are the basis to ensure the accuracy of maceral identification.However,limited by the low accuracy of acquisition equipment and non-standard manual operation,the actual collected coal photomicrographs have the problems of insufficient resolution and blur distortion,which cannot meet the requirements of maceral analysis tasks.Therefore,it is necessary to reconstruct the low-resolution image.Compared with traditional image reconstruction algorithms,the super-resolution reconstruction algorithm based on deep learning has the advantages of high accuracy and strong robustness.In view of this,this thesis aims to restore coal photomicrographs with high-fidelity and clear textures,and studies the super-resolution reconstruction method of coal photomicrographs based on deep learning.The main contents are as follows:Aiming at the problem of blurred details and poor visual perception in traditional convolutional neural network reconstruction images,a super resolution reconstruction model based on generative adversarial network was constructed.Firstly,a wide residual block with nested residual structure is designed in the generator.This block employs sub-residual structure to replace the batch normalization layer in the traditional residual block,which makes the color and contrast of the reconstructed image more closer to the real image and enhances the representation ability of the model for deep semantic information.In addition,the Mish activation function is employed in the discriminator to improve the nonlinear expression ability of the model.Experiment results show that the Peak Signal-to-Noise Ratio(PSNR)and Structural Similarity(SSIM)of the proposed generative adversarial network on the test dataset reach 30.6882 and 0.8997 respectively,which are superior to other classical super-resolution algorithms.The results of ablation experiment confirm the effectiveness of the wide residual block and the Mish activation function.Aiming at the problem that the existing model structure cannot fully utilize the correlation information between different feature layers,a Multi-scale Attention(MA)block is embedded in the generator of the model,which can capture non-local feature correlation information across scales.Comparative experiments were carried out for the location of the MA block and the number of layers of the feature pyramid in the block.The results show that the MA block with 5-layer feature pyramid structure embedded in the middle of the generator can achieve the highest PSNR and SSIM,which are 31.1210 and 0.9055,respectively.Aiming at the problem of lacking open source coal photomicrographs reconstruction tools,this thesis develops a software integrating the above models based on Py Qt5,which can realize the functions of image enhancement and super-resolution reconstruction,and provides convenience for researchers to obtain high-resolution coal photomicrographs that can be employed for research and analysis.This thesis has 49 figures,9 tables and 90 references.