Weakly supervised semantic segmentation of microscopic carbonates on marginal devices

Keran Li^a,1,*, Yujie Gao^b,1, Yingjie Ma^b,1, Chengkun Li^b, Junjie Ye^b, Hao Yu^b, Yiming Xu^b, Dongyu Zheng^c, Ardiansyah Koeshidayatullah^d

^aState Key Laboratory of Mineral Deposit Research, School of Earth Sciences and Engineering, Nanjing University, Nanjing 210023, China

^bCollege of Computer Science and Cyber Security, Chengdu university of Technology, Chengdu, 610059, China

^cState Key Laboratory of Oil and Gas Reservoir Geology and Exploitation and Key Laboratory of Deep-time Geography and Environment Reconstruction and Applications, MNR and Institute of Sedimentary Geology, Chengdu University of Technology, Chengdu 610059, China

^dDepartment of Geosciences, College of Petroleum Engineering and Geosciences, King Fahd University of Petroleum and Minerals, Dhahran 31261, Saudi Arabia

¹Equal contribution ^*Corresponding authors

📄 Paper | 🌐 Project Page

---

Fig 1. GIF performance of the model

Introduction

Our main tasks so far are to train a microscopic carbonate images classification model on a extremely large image datasets (Fig. 1). The datasets include over 10K+ high-resolution optical images. The images are mainly from practical hydrocarbon exploration projects in Sichuan basin. The datasets is named as Carbonate-Sichuan-170G from the location and volume of the datasets. 22 types of carbonate frameworks are carfully divided.

After training by a ResNet101 (parameters can be seen in Table 1, net framework can be seen in Fig. 2a), the .pth file to extract the corresponding heatmaps (CAM) on a lighter datasets, from Qi Z., Hou M., Xu S., et al., A microscopic image dataset of Sinian carbonate from Dengying Formation on the northwestern margin of Upper Yangtze. Science Data Bank, 2020. (2020-07-31). DOI: 10.11922/sciencedb.j00001.00105., shorten as "MidDynuy". The pre-trained ResNet101 took a role as a Teacher Net and distilled a light Student Net in MidDynuy (Fig. 3-4). The Student Net was a less than 800k ultra-lightweight model (original framework of MobileNetV3-Small can be seen in Fig. 2b; improved framework of MobileNetV3-Small can be seen in Fig. 3a; Details of improvements are displayed in Fig.3c-d). The 800k ultra-lightweight model adapted the MobileNetV3-Small framework.

Table 1 Parameters of ResNets

Fig 2. Original Model Frameworks of (a) ResNet101 and (b) MobileNetV3-Small

Fig 3. Frameworks improved MobileNetV3-Small; (a) Total framework; (b) Illustration of one bottleneck; (c) Flow of bottlenecks; (d) Illustration of the skip; (e) Flow of contran necks

To investigating the potential deploying scenario, in addition to CAM, other image processing such as glcm enhancement was used in the final rendering.

The workflow of this research is:

Fig 4. Workflow of this research

The effect and comparison of generating heatmaps are shown below

The following shows the original image to CAM heatmap to the GLCM enhanced CAM heatmap

Here are some more examples

Here's how to works in camera

Citation

If you use CamNet in your research, please cite:

@article{li2025weakly,
  title={Weakly supervised semantic segmentation of microscopic carbonates on marginal devices},
  author={Li, Keran and Gao, Yujie and Ma, Yingjie and Li, Chengkun and Ye, Junjie and Yu, Hao and Xu, Yiming and Zheng, Dongyu and Koeshidayatullah, Ardiansyah},
  journal={Computers \& Geosciences},
  pages={106059},
  year={2025},
  publisher={Elsevier}
}