基于深度学习的轨条砦障碍毁伤等级识别研究

Rail Obstacles Damage Level Recognition Based on Deep Learning

针对传统毁伤评估方法需要大量试验数据修正毁伤评估模型,且需要对炸点位置进行测量并结合弹目交汇条件才能计算毁伤效果的问题,提出了一种基于图像识别的端到端轨条砦障碍毁伤等级识别方法,进而快速对毁伤后轨条砦进行毁伤等级判定。该方案基于改进的YOLO5算法实现轨条砦图像检测、毁伤特征识别和毁伤等级评估。首先,由专家对轨条砦数据集进行标注,利用改进的马赛克拼接提高小目标检测能力;然后,使用卷积块注意力模块代替部分卷积模块,提升模型整体性能;最后,针对轨条砦的尺寸特点优化检测端结构,移除大目标尺度特征以减少推理计算量。实验结果表明:该方法可以在保证检测实时性的同时,毁伤评估精度达到98%。

In traditional damage assessment methods,the modification of the damage assessment model requires a large amount of experimental data,while the damage effect could not be calculated without the measurement of the location of the explosion point combined with the intersection conditions of the projectile and target.To solve the problem,an end-to-end rail obstacles damage level recognition method based on image recognition was proposed,by which the damage level of damaged rail obstacles could be determined quickly.Based on YOLO5 algorithm,the proposed method could realize rail obstacle images detection,damage characteristics recognition and damage level assessment.Firstly,the rail obstacle dataset was marked by experts,thereby the detection ability of small targets was improved by modified mosaic stitching.Then,convolutional block attention modules were used to replace some convolutional modules to improve the overall performance of the model.Finally,the detection end structure was optimized based on size characteristics of rail obstacles,and large target scale features were removed to reduce the load of the inference computation.Experimental results on the test set showed that the proposed method could realize real-time detection with a damage assessment accuracy of 98%.