Improvement of Mask R-CNN Algorithm for Ore Segmentation
Kai Tang, Yuguo Pei, Xiaobo Wang, Leilei Qu · Electronics · 2025
In response to the low precision of ore image segmentation under complex working conditions, an improved Mask R-CNN segmentation algorithm is proposed. The traditional Mask R-CNN uses a simple deconvolution operation to generate masks, which can lead to the loss of ore edge information and insufficient detail processing, affecting segmentation accuracy. Therefore, an improved model based on the Mask R-CNN framework is proposed in this paper. By introducing the Re-parameterized Refocus Convolution (RefConv) into the residual networks, the expressive power of the feature extraction network is enhanced. Meanwhile, the Efficient Channel Attention (ECA) is embedded in the output part of the Feature Pyramid Network (FPN), enhancing the model’s ability to capture key information. The improved Mask R-CNN network structure can reduce the loss of ore detail information caused by convolution operations and improve the network’s segmentation accuracy. Comparative experiments between the improved algorithm and the original algorithm show that the average Intersection over Union (MIoU) of the improved algorithm reached 92.8%, which is about a 6.8% increase compared to the original Mask R-CNN algorithm; the average pixel accuracy (mAP) is 97.2%, which is about a 5.1% increase compared to the original algorithm, indicating higher detection accuracy for ore identification and segmentation.