The Application Of Deep Learning In Mineral And Pores Recognition For The Rock Image

Traditional rock sample recognition technology relies on experts to examine and identify samples under a microscope.This process is time-consuming,labor-intensive,and challenging for sample preservation and sharing.In contrast,rock images,as a type of digital resource,have advantages such as easy preservation and high sharing.Based on this,digital rock identification can be further promoted towards the direction of automation and intelligence.In recent years,automatic rock image recognition method has become a research hotspot.Due to the difficulties of rock image with numerous targets and different scales,most scholars choose the strategy based on the combination of supervised learning and deep learning,and a lot of research results have been achieved.However,due to the lack of public labeling data in this field,manual labeling is extremely expensive,which brings great challenges to the further promotion and application of relevant research results.Addressing the constraints posed by limited labeled rock samples,this study employs deep learning and computer vision techniques to achieve automatic recognition of rock minerals and pores(component)using two distinct approaches.The two approaches are multi-step method based on feature analysis and deep learning method based on end-to-end.The former uses interpretable machine learning paradigms,and some processes require human intervention.The latter is based on a pure data-driven paradigm with a higher degree of automation.By comparing and analyzing these methods,we provide more reliable support for further technological selection and contribute to the development of explainable deep learning models.The primary research content and innovations of this study include:(1)A multi-step rock image component recognition method based on feature analysis.This method divides multi-component rock image recognition into three steps: presegmentation,region merging,and single-component rock image recognition.We initially employ the simple linear iterative clustering(SLIC)algorithm for rock image pre-segmentation,followed by region merging based on color features and cropping using the minimum bounding rectangle,and only 6% of the cut single-component rock dataset needed to be manually labeled.Then use this data to continuously generate a highly reliable pseudo-label by an improved semisupervised dual model training method.This approach yields a single-component rock classification model with 96.3% accuracy,which we extend to the entire dataset.Finally,we use the SLIC algorithm and the region merging algorithm to generate sub-regions for rock images,inputting these sub-regions into the single-component rock classification model to obtain rock image component recognition results.Our method demonstrates high generalization and reliability,significantly reducing data labeling efforts and offering a reference for research in other fields with limited labeled data.(2)A rock image component recognition method based on Mask R-CNN.In order to further improve the degree of automation of the above methods,using the Mask R-CNN model as a benchmark and limited rock label data as the foundation,we first train an initial rock component recognition model through supervised learning.We then optimize the recognition results using three methods: Method I involves initializing the Mask R-CNN model with existing weights,leading to a 21.7% improvement in mean Average Precision(m AP)and accelerated learning efficiency compared to the initial model.Method II proposes an improved weakly supervised learning algorithm,supervised-like learning,based on the existing singlecomponent rock label data,generating a large amount of weak label data containing complete label information and mitigating the impact of noise on the original label data.Combining this approach with existing weight initialization models and weak label data,we achieve a 28.2%improvement in m AP compared to the initial model.Method III designs a semi-supervised algorithm based on the SE-Res Net18 Discriminator to effectively utilize a large amount of unlabeled rock data.This algorithm measures the reliability of pseudo-labels using the discriminator,continually iterating and optimizing the labeled data and Mask R-CNN model,and resulting in a 45.5% increase in m AP compared to the initial model.Our analysis of actual data testing demonstrates that these three optimization methods effectively improve rock image recognition results,addressing the challenges posed by the particularity of rock images and insufficient labeled data.