Geological map feature extraction using object detection techniques - a comparative analysis
| Autor: | P. A. N. Dilhan, R. Siyambalapitiya |
|---|---|
| Rok vydání: | 2022 |
| Předmět: | |
| Zdroj: | i-manager’s Journal on Image Processing. 9:28 |
| ISSN: | 2349-6827 2349-4530 |
| DOI: | 10.26634/jip.9.2.18916 |
| Popis: | Conducting a geological field survey at the initial stage is an important step in geo-oriented projects and construction. Therefore, better and more accurate solutions are only possible with field analysis and proper modeling. The geological modeling process takes a long time, especially depending on the area of interest. It is inefficient to digitize 2D geological maps with traditional software that uses manual user interaction. This paper proposes a state-of-the-art feature detection methodology for detecting geological features on high-resolution maps. With the development of efficient deep learning algorithms and the improvement of hardware systems, the accuracy of detecting specific objects in digital images, such as human facial features, has reached more than 90%. Current object detection models based on convolutional neural networks cannot be directly applied to high-resolution geological maps due to the input image size limitations of conventional object detection solutions, mostly limited by hardware resources. This paper proposes a sliding window method for character detection of geological features. Detection models are trained using transfer learning with You Look Only Once-v3 (YOLO-v3), Single Shot Multi-Box Detector (SSD), Faster-Region-based Convolutional Neural Network (Faster-RCNN), and Single Shot Multi-Box Detector_RetinaNet (SSD_RetinaNet). All models provide competitive success rates with an average precision (AP) of 0.96 on YOLOv3, 0.88 AP on EfficientNet, 0.92 AP on Faster- RCNN, and 0.97 AP on SSD_RetinaNet. YOLOv3 outperformed the best detection over SSD according to F1 recall and score. Since the input size of detection models is limited, a sliding window algorithm is used to separate high-resolution map images. The final detected strike features are provided as a digital dataset that can be used for further manipulations. Thus, Convolutional Neural Network (CNN) based object detection along with a sliding window protocol can be applied to manual map digitization processes to provide instantaneous digitized data with higher accuracy. This automated process can be used to detect small features and digitize other high-resolution drawings. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|