SCAR-UNet: an Improved Res-UNet with Channel and Spatial Attention for Coal Maceral Image Segmentation

EasyChair Preprint 14670

Abstract

Coal, as a versatile natural resource fueling various industries, necessitates accurate identification of its maceral components for mining and geological applications. However, automated segmentation of coal macerals remains challenging due to the grayscale similarity between maceral components like liptinite and the background in coal photomicrographs. This study proposes SCAR-UNet, a novel improved network architecture designed specifically for maceral image segmentation. Our approach integrates channel attention, spatial attention, and a novel loss function with the Residual UNet (Res-UNet) architecture, enabling enhanced feature extraction and model performance. Evaluated on a labeled Coal Maceral image dataset comprising 908 annotated images containing vitrinite, inertinite, and liptinite macerals. The widely used Intersection over Union (IoU) and Pixel Accuracy (PA) metrics were utilized for assessing segmentation performance. The proposed SCAR-UNet outperformed state-of-the-art segmentation algorithms.

Keyphrases: Attention Mechanism, Coal Maceral Segmentation, Residual UNet, bdl loss function, channel attention and spatial attention, coal rock microscopic images, deep learning, image segmentation, loss function, microscopic image segmentation, segmentation of coal, semantic segmentation models

@booklet{EasyChair:14670,
  author    = {Fanqian Meng and Ouli Luo and Hui Ding and Guoping Huo},
  title     = {SCAR-UNet: an Improved Res-UNet with Channel and Spatial Attention for Coal Maceral Image Segmentation},
  howpublished = {EasyChair Preprint 14670},
  year      = {EasyChair, 2024}}