Research On Automatic Recognition Method Of Tectonic In Field Geological Outcrop Image

Structural feature recognition is of great significance for geological disaster investigation,urban planning,and seismic risk assessment.The traditional preliminary geological survey methods based on field reconnaissance have great risks in high and steep areas.With the widespread application of technologies such as unmanned aerial vehicle tilt photography,a large number of geological outcrop images can be obtained.The existing image-based structural feature recognition requires manual interpretation for structural positioning and classification.Manual interpretation has problems such as high intensity,high cost,and strong subjectivity.Because the rapidly develop of deep learning in every aspect,it provides an important foundation for intelligent structural feature recognition.There are the following two types of problems in the extraction of structural features of geological outcrop images in the wild.First of all,in image data processing,there are problems such as cumbersome data preprocessing,multiple solutions of feature annotation forms,and insufficient data training samples.In the structure feature recognition,there are problems of low manual interpretation efficiency and insufficient automatic recognition accuracy.In response to the above problems,this paper proposes an automatic recognition method for the structural features of the geological outcrop image based on deep learning.Using Mask R-CNN instance segmentation network,Canny edge detection algorithm,Fully Convolutional Networks(FCN)and other deep learning methods,a neural network model for rapid geological structure feature classification and recognition is constructed.This subject is mainly researched from three aspects:field geological outcrop image structure features and collection methods,data set production methods and automatic identification methods of structure features.Aiming at the problem of automatic batch acquisition of field geological outcrop images,this paper obtains high-resolution field geological outcrop images from the Internet and literature to prepare the data for the training set.In view of the large difference in image data format from different sources,this paper proposes the method of field geological outcrop images.The normalization process also uses rotation,blur,and grayscale methods for image augmentation to expand the number of training samples and provide a sufficient data basis for the identification of geological outcrop structure features in the field.Aiming at the problem of automatic interpretation of structural features,this paper uses the Mask R-CNN network based on bilinear difference improvement,and uses Res Net-101 as the backbone network to extract the structural features in the image,so as to realize the intelligent target detection and positioning of the structural features in the geological outcrop image.Based on the methods provided above,this paper conducted experiments with obvious faults,folds,and stone sausage structural features.The results show that the normalization process and preprocessing method of image data from different sources proposed in this topic can automatically generate batches for model training.The image preprocessing method effectively expands the training sample library without destroying the structural features of the image.This paper is based on the improved Mask R-CNN instance segmentation method to effectively realize the classification and recognition of structural features in the geological outcrop images in the field.