Weakly supervised semantic segmentation of microscopic carbonates on marginal devices

Microscopic analysis is the cornerstone to uncover petrological and mineralogical characteristics of carbonate rocks. In addition, such information is critical for precise identification of carbonate microfacies and diagenetic evolution. This type of information is important, but relies too much on manual experience, which is time-consuming and laborious. Recently, several successful deep learning models showed great potential in the identification process. However, current deep learning models have typically complex model architectures greatly hinder the deployment-inference in practical and lightweight environments. To overcome the difficulty of deep learning models in reasoning in actual edge scenes, a three-stage segmentation method by weakly supervised learning was proposed. The approach embeds class activation mapping (CAM), grey level co-occurrence matrix (GLCM), and knowledge distillation (KD) modules to achieve attention transfer to the lightweight network (CamNet). Furthermore, based on the performance of the model algorithm and application requirements, a lightweight carbonate thin section image-assistant recognition system has been developed. Through ingenious control flow design, this system achieves an effective balance between runtime latency and resource consumption, demonstrating superior performance metrics. Experimental results indicate that CamNet's total parameter count is only 800k. When deployed in embedded systems, CamNet achieves an inference speed of 6.87 fps. Our successful development verifies the efficiency and practicality in marginal devices.