浮选图像中泡沫边缘分割  

Froth Edge Segmentation in Flotation Image

作　　者：刘文卓 乔亦诚 殷雪原 刘晶 甘咏祺 李宗泽 张国英[1] LIU Wenzhuo;QIAO Yicheng;YIN Xueyuan;LIU Jing;GAN Yongqi;LI Zongze;ZHANG Guoying(School of Artificial Intelligence,China University of Mining&.Technology Beijing,Beijing 100083,China;School of Vehicle and Mobility,Tsinghua University,Beijing 100084,China)

机构地区：[1]中国矿业大学(北京)人工智能学院,北京100083 [2]清华大学车辆与运载学院,北京100084

出　　处：《有色金属（选矿部分）》2024年第4期67-76,共10页Nonferrous Metals（Mineral Processing Section）

基　　金：国家自然科学基金资助项目(U1704242)。

摘　　要：矿物浮选过程中泡沫的表面特征是浮选性能的重要指标,它可以实时、直观地反映浮选效果的变化,准确分割泡沫的边缘信息是浮选过程中一项重要的任务。近年来,研究人员提出了各种浮选泡沫图像分割算法,但浮选泡沫图像中存在泡沫数量多、泡沫间粘连严重以及边缘不清晰等问题,现存的方法由于其特征提取能力有限,无法精确地分割泡沫边缘。据此,利用深度学习提出了一种基于多尺度融合的浮选泡沫图像边缘分割算法,该算法通过引入一种深度高分辨率的编码结构以及一种基于注意力的分层融合方法来增强模型的特征提取能力,从而提高对于浮选泡沫边缘的分割效果。具体而言,深度高分辨率的编码结构可以在不同分辨率层级上同时维护特征信息,使我们的网络模型可以有效地捕捉不同尺度的信息,在提高图像语义理解能力的同时能够保持更多的细节信息,提高处理高分辨率以及密集任务图像的能力。除此之外,设计了一种基于注意力的分层融合方法来充分融合深层和浅层的特征图,使融合得到的特征图趋向于更重要的特征信息,从而提高识别浮选泡沫的边界和精确定位浮选泡沫的能力。该算法在泡沫边界分割数据集上凭借58.25的泡沫IoU以及73.62的泡沫F_(score)取得了最佳的性能,证明了我们提出的算法可以更加准确地分割浮选泡沫边缘。The surface characteristics of the froth in the mineral flotation process is an important indicator of flotation performance,which can reflect the change of flotation effect in real time and intuitively,and accurate segmentation of the edge information of the froth is an important task in the flotation process.In recent years,researchers have proposed various flotation froth image segmentation algorithms,but there are problems such as large number of froths,serious adhesion between froths and unclear edges in the flotation froth image,and the existing methods are unable to accurately segment froth edges due to the limited feature extraction capability.Accordingly,a multi-scale fusion-based edge segmentation algorithm for flotation froth images using deep learning has been proposed,which enhances the feature extraction capability of the model by introducing a deep high-resolution coding structure and an attention-based hierarchical fusion method to improve the segmentation effect of flotation froth edges.Specifically,the deep high-resolution coding structure can simultaneously maintain feature information at different resolution levels,so that our network model can effectively capture information at different scales,improve image semantic comprehension while maintaining more detailed information,and improve the ability to process high-resolution and task-intensive images.In addition to this,an attentionbased hierarchical fusion method has been designed to fully fuse the deep and shallow feature maps,so that the feature maps obtained from the fusion converge to more important feature information,thus improving the ability to recognize the boundaries of flotation froths and accurately locate flotation froths.The algorithm achieves the best performance on the froth boundary segmentation dataset with a froth IoU of 58.25 and a froth F_(score) of 73.62,proving that our proposed algorithm can segment flotation froth edges more accurately.

关 键 词：浮选泡沫 图像分割 深度学习 卷积神经网络 

分 类 号：TD923.7[矿业工程—选矿]