Improved ResUNet for Semantic Segmentation of Complex Multi-Scale Graphite Ore Images
Year: 2026
Type: article
Abstract: To address the challenging task of precise segmentation in graphite mineral images, where large flake graphite, discrete fine particles, and embedded micro-particles coexist alongside strong background noise interference from surrounding rock, an improved ResUNet semantic segmentation model named WRAG-ResUNet is proposed. This model employs ResNet50 as the encoder backbone, replacing the initial max pooling layer with a wavelet-inspired downsampling module to preserve high-frequency edge information of fine graphite particles during downsampling. At the deepest encoder layer, a dilated spatial pyramid pooling module is embedded, utilizing parallel multi-group dilated convolutions with dilation rates of 6, 12, and 18 to extract multi-scale contextual features suitable for graphite targets of varying sizes. A residual attention gate module with skip connections is introduced in the skip connection path, adaptively weighting and filtering shallow encoder features using high-level decoder semantic features to enhance target features and distinguish them from the background while retaining effective details. During training, a combined loss function of Focal Loss and Dice Loss is employed to focus on hard-to-classify boundary pixels and small discrete fragments. Experimental results on a self-built graphite mineral image dataset demonstrate that WRAG-ResUNet achieves pixel accuracy and mean intersection over union (IoU) rates of 96.25% and 92.95%, respectively, representing increases of 3.88 percentage points and 5.49 percentage points compared to the classic U-Net. The model’s comprehensive performance surpasses mainstream generic semantic segmentation models such as FCN-8s, DeepLabV3+, ResUNet, and U-Net++, effectively enhancing segmentation accuracy for graphite targets of diverse morphologies and providing a preliminary technical reference for laboratory-scale analysis of ore images under specific conditions.
Source: Applied Sciences
DOI: 10.3390/app162211198
Authors: Mengguang Yang, Xueyu Huang, Le Chen, Binghui Wei
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,896
OpenAlex ID: W7225878519
