At present, there are some problems in the process of gravity separation shaking table beneficiation, such as low efficiency and high cost of manual identification of ore belts, as well as large number of parameters and insufficient accuracy of existing segmentation models. This study addresses these challenges by developing a model that is both computationally efficient and highly accurate for the segmentation of scheelite zones under ultraviolet light, a condition that enhances the natural fluorescence of the target mineral. Central to this research is a significantly enhanced DeepLabV3+ model, which integrates several critical architectural modifications. Initially, the computationally intensive Xception backbone was substituted with the more efficient MobileNetV3 network, which was further enhanced by integrating an efficient channel attention module. This modification aimed to reduce the parameter count while enhancing feature representation. Subsequently, the encoder’s atrous spatial pyramid pooling module was restructured into a DenseASPP configuration. This redesign incorporated additional atrous convolutional branches to establish a more continuous range of receptive fields, thereby facilitating superior multi-scale feature extraction from mineral bands of varying widths. Furthermore, the decoder was augmented with a convolutional block attention module to highlight discriminative features and suppress noise, complemented by deformable convolutions to accommodate the irregular shapes and boundaries of the mineral zones. The model was trained and evaluated on a custom-developed dataset of ultraviolet-lit scheelite images, which was expanded using data augmentation techniques. A series of ablation studies confirmed the contribution of each proposed enhancement. In comparative experiments, the final model exhibited superior performance, achieving a Dice coefficient of 93.89%, a mean intersection over union of 91.97%, and an mAP@50:95 of 81.04%. The study demonstrates that the proposed model achieves a high level of accuracy with a notably reduced model size of 18.32 million parameters and a computational cost of 33.50 GFLOPs. This indicates an optimal balance between precision and efficiency, making it well-suited for industrial applications. The research conclusively shows that the model effectively fulfills the dual requirements of high accuracy and a lightweight architecture for segmenting shaking table mineral zones. By incorporating specific advancements in network architecture, the model offers a robust technical foundation for automating shaking table control systems. This development has the potential to significantly improve beneficiation efficiency and consistency, marking a substantial advancement in intelligent mineral processing. Future research will focus on further validating the model’s generalization capabilities across a broader range of industrial operating scenarios.
国内外学者围绕摇床矿带的视觉分割开展了诸多研究。摇床分选过程的智能化研究,其技术范式经历了从机理模型到数据感知的演进。早期研究侧重于从物理原理出发进行理论推导和仿真,如:通过流体流动模拟来建立增强型摇床的配分函数模型(Kroll-Rabotin et al,2010)。然而,由于实际生产环境复杂多变,基于直接感知的视觉检测与数据驱动方法因其更强的适应性而成为近年来的主要研究方向,并沿着“图像处理—智能识别—系统集成”的路径快速发展。在图像处理与特征提取阶段,研究人员采用多种算法从图像中直接定位矿物带边界,例如:融合深度学习边缘检测与Hough变换,提取了摇床矿带的重要边界(刘惠中等,2023);采用MSR进行了图像预处理,并使用OTSU算法进行左侧矿带图像的分割(王锦煜,2022),还有磷虾优化算法(和丽芳等,2018)以及利用最小二乘法获得摇床分割带中心矢量的方法(杨文旺等,2020)。随着科技的进步,研究重点进入深度学习智能预测阶段,旨在建立端到端的映射模型,如:使用VGG-16网络预测接矿位置(赵玉华等,2023),设计更为复杂的CNN-GRU-Attention混合模型进行精矿边界的回归定位(刘惠中等,2025)。部分研究正逐步向系统集成和自动控制工程化深入,如:杨文龙等(2016)开发的集识别和调节于一体的自动选矿装置,标志着该项技术向闭环控制与工业落地迈出了关键一步。
DeepLabV3+模型是在DeepLabV3基础上进行的重要升级(Chen et al,2018),后者作为语义分割领域的代表性模型,已被广泛应用于各类语义分割任务中。DeepLabV3+模型在编码器部分延续并增强了空洞空间金字塔池化(ASPP)模块,通过多个不同空洞率的并行卷积和全局池化操作,有效融合多尺度上下文语义信息。在解码器设计中,模型引入轻量级上采样结构,将经过通道压缩的低层特征与富含语义的高层特征进行拼接融合,显著提升了目标边界的重建精度。为进一步实现效率与性能的平衡,DeepLabV3+模型支持通过调整输出步幅和相应层的空洞率,灵活控制感受野大小和计算复杂度。在损失函数方面,除了使用交叉熵作为基础监督,还可结合Dice损失或针对性设计的边界感知损失。DeepLabV3+模型的具体结构如图1所示。
(1)骨干网络改进:针对原有DeepLabV3+模型骨干网络参数量大和特征提取不足的问题,本研究使用添加了ECA(Wang et al,2020)注意力机制的MobileNetV3骨干网络(Howard et al,2019)代替原有的Xception骨干网络(Chollet,2017)。MobileNetV3的轻量化特性使得模型分割精度在小幅下降的同时,极大地降低了模型的计算量,同时ECA增加了模型的表征能力和浅层特征提取能力,在降低参数量的同时提升了模型对矿带的浅层能力,弥补了替换骨干网络后导致的分割精度的缺失。
(3)解码器改进:在解码器中添加可变形卷积(Xia et al,2022)和CBAM(Woo et al,2018)注意力机制。可变形卷积通过为标准卷积核的每个采样点引入一个可学习的偏移量,使卷积核能够根据输入特征的内容自适应地调整其感受野的形状,从而在矿带形变时,能够更好地拟合矿带的形状,从而提升分割精度。CBAM的通道注意力机制放大了不同矿带之间的微小差异,强化了矿带的纹理特征和颜色特征,抑制背景噪声,提升相近矿带的区分度,最终让解码器输出更精细的分割结果。
本研究使用ECA模块替代MobileNetV3自带的SE模块(Hu et al,2018),ECA模块具体结构如图2所示。ECA模块通过全局平均池化获取通道特征的全局描述,并摒弃了传统的全连接层,转而采用一维卷积来直接计算通道间的局部交互,从而生成通道权重。这种方法在避免降维带来的信息损失的同时,以极低的计算开销实现了跨通道的信息交互,引导模型聚焦于关键特征通道。
为了解决多尺度矿带特征捕获不连续的问题,本研究对DeepLabV3+模型中的ASPP模块进行了改进。传统的ASPP模块采用多分支并行结构,其各分支的感受野离散(DeepLabV3+自带的池化金字塔膨胀率为r=6,12,18),且分支间无信息交互,导致对不同宽度矿带的特征提取存在间隙,这是一种“迟融合”策略。为进一步增强模型对从细矿带到宽矿带的连续尺度特征的捕捉能力,本文设计了一种密集连接的空洞卷积(Yu et al,2016)金字塔池化模块,其结构如图3所示。
为了验证Ours模型在紫外灯下摇床矿带分割任务中的性能表现,开展了系统的对比试验。分别使用U-NET(Ronneberger et al,2015)、U-NET++(Zhou et al,2018)、HRNet(Wang et al,2021)、PSPNet(Zhao et al,2017)、DeepLabV3+和Ours共6种网络模型进行训练和评估。在试验过程中,所有模型均在相同的硬件环境下运行,且采用相同的训练数据集、迭代次数和超参数设置,以确保试验结果的公平性和可比性。图13所示为不同模型的分割效果图。其中,红色、绿色和黄色区域分别为精矿带、中矿带和尾矿带的语义分割结果。
ChenL C, ZhuY K, PapandreouG,et al,2018.Encoder-decoder with atrous separable convolution for semantic image segmentation[C]//Computer Vision-ECCV 2018.Cham:Springer,833-851.
[2]
CholletF,2017.Xception:deep learning with depthwise separable convolutions[C]//2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR).Honolulu:IEEE,1800-1807.
[3]
HowardA, SandlerM, ChenB,et al,2019.Searching for MobileNetV3[C]//2019 IEEE/CVF International Conference on Computer Vision(ICCV).Seoul:IEEE,1314-1324.
[4]
HuJ, ShenL, SunG,2018.Squeeze-and-excitation networks[C]//2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition.Salt Lake:IEEE,7132-7141.
[5]
Kroll-RabotinJ S, BourgeoisF, ClimentÉ,2010.Fluid dynamics based modelling of the Falcon concentrator for ultrafine particle beneficiation[J].Minerals Engineering,23(4):313-320.
[6]
RonnebergerO, FischerP, BroxT,2015.U-Net:convolutional networks for biomedical image segmentation[C]//Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015.Cham:Springer,234-241.
[7]
WangJ D, SunK, ChengT H,et al,2021.Deep high-resolution representation learning for visual recognition[J].IEEE Transactions on Pattern Analysis and Machine Intelligence,43(10):3349-3364.
[8]
WangQ L, WuB G, ZhuP F,et al,2020.ECA-net:efficient channel attention for deep convolutional neural networks[C]//2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Seattle:IEEE,11531-11539.
XiaZ F, PanX R, SongS J,et al,2022.Vision transformer with deformable attention[C]//2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR).New Orleans:IEEE,4784-4793.
[11]
YuF, KoltunV,2016.Multi-scale context aggregation by dilated convolutions[C]//Bengio Yoshua,LeCun Yann.4th International Conference on Learning Representations (ICLR) 2016.JuanSan,Puerto Rico:International Conference on Learning Representations.
[12]
ZhaoH S, ShiJ P, QiX J,et al,2017.Pyramid scene parsing network[C]//2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR).Honolulu:IEEE,6230-6239.
[13]
ZhouZ W, Rahman SiddiqueeM M, TajbakhshN,et al,2018.UNet++:a nested U-Net architecture for medical image segmentation[C]//Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support.Cham:Springer,3-11.
[14]
和丽芳,郭思哲,黄宋魏,等,2018.一种基于磷虾优化算法的摇床矿带分带图像分割法:CN105844628B[P].2018-06-15.He Lifang,Guo Sizhe, Huang Songwei, et al,2018. An image segmentation method based on krill optimization algorithm for zoning of rock bed ore zone: CN105844628B[P]. 2018 06 15.
LiuHuizhong, LiuJianye, HuangAo,et al,2025.Research on the construction of the relationship model of“control parameter-concentrate boundary coordinates”in shaking table[J].Gold Science and Technology,33(3):648-658.
LiuHuizhong, YouKeshun,2023.Boundary extraction and feature representation of rocking mineral belt through deep learning edge detection and Hough transform[J].Manufacturing Automation,45(6):180-183,188.
LiuXianyu, YangJiuchang, TuCai,et al,2023.Study on spectral characteristics of scheelite from Xuebaoding,Pingwu County,Sichuan Province,China[J].Spectroscopy and Spectral Analysis,43(8):2550-2556.
TianNiu, HuangSongwei, HeLifang,et al,2023.Application of deep learning in mineral processing field[J].Industrial Minerals & Processing,52(8):69-74,82.
[23]
王锦煜,2022.摇床矿带识别算法及其应用[D].湘潭:湖南科技大学.Wang Jinyu,2022.Algorithm and application of shaking ore belt identification[D].Xiangtan:Hunan University of Science and Technology.
[24]
杨文龙,吴富姬,罗璇,等,2016.一种选矿摇床矿带识别及接矿板自动调节装置:CN105903557A[P].2016-08-31.Yang Wenlong,Wu Fuji,Luo Xuan,et al,2016.An ore belt identification and automatic adjustment device for ore receiving plate of mineral processing shaking table:CN105903557A[P]. 2016 08 31.
[25]
杨文旺,何庆浪,兰希雄,等,2020.选矿摇床智能巡检机器人开发与应用[J].有色金属(选矿部分),(5):102-106.Yang Wenwang,He Qinglang,Lan Xixiong,et al,2020.Development and application of intelligent inspection robot for table concentrator[J].Nonferrous Metals Mieral Processing Section,(5):102-106.
[26]
赵玉华,杨文旺,武涛,等,2023.基于深度学习的摇床接矿位置预测方法[J].有色金属(选矿部分),(2):142-146,172.Zhao Yuhua,Yang Wenwang,Wu Tao,et al,2023.Prediction method of shaking table contacting position based on deep learning[J].Nonferrous Metals (Mineral Processing Section),(2):142-146,172.
ZhengYurong, JinJunbao, WuXinnian,et al,2025.Research and application progress of key technologies in smart mine[J].Gold Science and Technology,33(2):404-416.