Research On Rock Image Segmentation In Wild Outcrops Based On Deep Learning

In geological survey,the rock image of outcrop area can provide key information for the study of oil and gas migration in strata.Therefore,it is a very important subject in the field of geology to study the fractures and rock types of rock images in outcrop areas.The traditional method of rock image segmentation in the field outcrop area mostly adopts the method of manual description,which is the most commonly used method in geological investigation.However,this method is inefficient,time-consuming and laborious,and greatly affected by the experience of observers,so it is difficult to provide quantitative data for analysis.With the development of image processing technology,the study of rock image segmentation in outcrop area eliminates the influence of human subjective factors and is more accurate than manual description.However,the traditional technology is still unable to provide high-precision data support for geological investigation and resource exploration models.In recent years,crack detection and segmentation methods based on machine vision have been widely developed,but these methods can only detect and segment simple cracks with high accuracy in a specific environment,and can only detect the effective information of cracks in this specific environment image,so it is difficult to classify cracks and other elements at the same time.In the actual outcrop environment,rock image segmentation is easily affected by environmental factors,which makes the results obtained by traditional methods have large errors and low generalization ability.In recent years,deep learning has developed rapidly in the field of image processing.Aiming at the rock image data set of outcrop area collected in the real environment,this paper proposes a rock image segmentation algorithm RC-Seg Net based on deep learning image semantic segmentation.The algorithm adopts the main structure of encoder-decoder.The encoder part is based on Res Net50 model to extract different levels of features,The CRA(collaborative refinement attention block)module is introduced to enhance the effective extraction of global context information of features,while retaining accurate location information,which helps the network locate the objects of interest more accurately.In addition,the low-level surface feature mapping of the encoder and the high-dimensional abstract feature mapping of the decoder are fused through long and short jump connections,so as to achieve more accurate segmentation of cracks and rock types in rock images.The results show that the image semantic segmentation algorithm model proposed in this paper achieves an overall recognition accuracy of 91.7% on the two indicators of mean accuracy and MIo U value,of which the detection accuracy of rock fractures reaches 87.8%and the recognition accuracy of rock types reaches 96%,which are significantly better than the classical semantic segmentation algorithms such as Seg Net,U-net,FPN,and Res Net50.The application of this method effectively improves the segmentation accuracy of rock images.