一种基于深度学习的X射线透射铀矿识别算法

A Uranium Ore Recognition Algorithm Based on Deep Learning of X-ray Transmission

铀资源是我国重要的战略资源,在工农业生产、科学技术以及国防领域中有广泛的用途,但我国已探明铀矿床主要以中低品位为主,矿体形态复杂,不连续分散分布,开采成本高,利用率低。为了识别可冶炼铀矿石,提高资源利用率,更好地支持国家的铀矿资源储备,采用X射线透射成像技术,并以收集到的铀矿石透射图像为研究对象,提出了一种名为SAxVit的轻量化高性能铀矿识别算法。该算法采用了ShuffleNet的思想,设计了一个轻量级的主干网络用于初步的特征提取。其次,提出了一种轴向平均注意力机制来进行深度特征提取,该机制通过计算特征图横轴纵轴上的平均值获取特征分布趋势矩阵,然后加权该矩阵计算特征图上每个像素点与其他像素点之间的关联性,以符合透射图像根据不同灰度值区分元素的工作原理。最后,基于Vision Transformer模型,剔除特征图分割操作,设计了一种即插即用的轻量化的ViT模块用作轴向平均注意力机制的载体。试验结果表明,SAxVit算法的参数量低至0.077M,识别速度低至3.481 ms,在铀矿透射图像测试集上该算法的识别准确率达到了95.7%。相比于MobileNetV2、ShuffleNetV2_0.5、Vit_64x64以及MobileVit_small,SAxVit在识别准确率、轻量化和识别速度等方面都取得了显著的改进。

Uranium resources are important strategic resources in our country,with wide-ranging applications in industrial and agricultural production,scientific and technological fields,as well as national defense.However,the identified uranium deposits in our country are mainly of medium to low grade,with complex ore body forms and discontinuous and scattered distribution.The mining costs are high,and the utilization rate is low.To identify uranium ores suitable for smelting,improve resource utilization,and better support the national uranium reserves,the X-ray transmission imaging technology is used,and the collected uranium ore transmission images are taken as the research object,and a lightweight highperformance uranium ore identification algorithm named SAxVit is proposed.The algorithm incorporates the principles of ShuffleNet and designs a lightweight backbone network for initial feature extraction.Additionally,a novel Axial-mean Attention mechanism is introduced for deep feature extraction.This mechanism calculates the average values along the horizontal and vertical axes of the feature map to obtain a matrix representing the distribution trend of features.The matrix is then weighted to compute the correlation between each pixel and other pixels on the feature map,aligning with the working principle of transmission imaging that distinguishes elements based on different grayscale values.Lastly,based on the Vision Transformer model,a lightweight plug-and-play ViT module is designed to serve as the carrier of the axial average Attention mechanism.The experimental results demonstrate that the SAxVit algorithm has a parameter count as low as 0.077M and achieves recognition speeds as fast as 3.481 ms.The algorithm achieves a recognition accuracy of 95.7%on the test dataset of uranium transmission images.In comparison to MobileNetV2,ShuffleNetV2_0.5,Vit_64x64,and MobileVit_small,SAxVit exhibits significant improvements in terms of identification accuracy,lightweight design,and recognition speed.