An ore image segmentation method based on an improved UNet model

Ore image segmentation is a critical step in determining the particle size distribution of ores, which imposes high demands on segmentation accuracy. However, challenges arise due to the diversity in ore sizes and shapes, the complexity of surface textures, uneven illumination, and ore adhesion. This paper proposes a mineral image segmentation method based on an improved the UNet model to address these issues. First, preprocessing techniques, such as grayscale conversion, histogram equalization, homomorphic filtering, and fixed-window cropping, are employed to enhance image quality and expand the dataset. Next, the UNet model structure is optimized using a lightweight design, reducing the number of feature channels by half to decrease computational complexity. The Tversky loss function is also integrated to improve segmentation performance in ore boundaries and dark regions. Experimental results demonstrate that the proposed method outperforms comparative models in terms of segmentation accuracy (e.g., Balanced Accuracy) and boundary refinement, providing reliable technical support for ore particle size analysis.