Research on Lightweight Annotation Method for Fragmented Ore Images Based on Improved U2Net

• Luo Xiaoyan,
• Wang Mengzhen,
• Gong Hanbiao,
• Luo Hui,

Vol. 20, No. 3, March 31, 2026
10.3837/tiis.2026.03.015, Download Paper (Free):

• Fragmented ore
• image annotation
• Automatic Annotation
• U2Net
• Lightweight

Abstract

To address the issues of high time consumption, labor intensity, quality instability, and low efficiency in manual annotation of ore images, this study proposes an improved lightweight automatic annotation model based on the U2Net network. Firstly, the Squeeze-and-Excitation (SE) module was introduced into the encoder stages to enhance global context modeling and mitigate the loss of feature information during sampling. Subsequently, the Convolutional Block Attention Module (CBAM) was introduced to further extract fine-grained features of fragmented ores and improve model performance. Based on the improved segmentation network U2Net++, this study constructs an automatic annotation network. By incorporating the Beetle Antennae Search (BAS) algorithm for parameter optimization and the Douglas-Peucker algorithm for contour simplification, we achieve lightweight automatic annotation of ore images. Experimental results show that the improved U2Net++ model achieves enhanced performance in fragmented ore image segmentation tasks, with a 4.19% increase in accuracy, 19.15% reduction in Mean Absolute Error (MAE), 2.13% improvement in precision, 2.21% boost in recall rate, and 1.96% enhancement in F1-score. The improved lightweight algorithm optimizes contour points while maintaining edge accuracy, significantly reducing JSON file size and demonstrating high consistency in Cohen's Kappa coefficient. This lightweight algorithm remarkably enhances ore annotation efficiency, providing crucial technical support for intelligent sorting.

Statistics

Cite this article