ELC U-Mamba: Efficient Linear-Chunked U-Mamba for Long-Range and Local Feature Fusion

The segmentation of coal microscopic component images is crucial for coal quality assessment and analysis, calling for automated approaches that can significantly enhance the reliability and speed of coal inspection processes. Due to the lack of open-source datasets for coal microscopic component images, research on deep learning algorithms in this field has been limited. To address these challenges, this paper proposed a novel ELC U-Mamba model, which integrates the ELC-MLA mechanism with the U-Mamba architecture. ELC-MLA reduces computational complexity and significantly enhances the efficiency of long sequence processing compared to traditional MLA mechanisms by employing linearized attention calculations and a dynamic chunking strategy. This model optimizes performance on complex tasks by effectively merging long-range and local features. Additionally, the Bridge module within the U-shaped structure combines ELC-MLA with the Mamba layer, achieving a balance between global information transmission and local information capture. To validate our approach, we constructed a specialized dataset comprising 1,684 coal microscopic images. Experimental results demonstrate that our method achieves superior performance on key metrics such as PA, IoU, and Dice, outperforming existing methods in the segmentation of coal microscopic component images.