摘要
基于图像的表面锈蚀检测是起重机损伤检测的重要新兴技术。由于起重机械的结构形状复杂,无人机等设备获取的起重机图像会包含大量的阴影以及黑色背景物,易被误识别为锈蚀;同时,起重机属于高端装备,锈蚀损伤样本量往往不足,加重了锈蚀检测的难度。针对以上问题,文中提出了一种基于二阶段网络的起重机小样本图像锈蚀检测算法。具体地,设计了一个包含分割和决策的二阶端到端学习网络,其中分割网络使用锈蚀损伤的标签进行训练,决策网络则根据分割网络提取的特征,进一步学习判断输入的图片中是否包含锈蚀损伤,并通过将二阶段的网络由类似于VGGNet的结构提升为类似于ResNet的残差结构来对网络进行优化。实验结果表明,提出的方法有效提升了图像锈蚀检测的精度,在自建的起重机小样本图像锈蚀数据集上,将模型的平均精度由94.2%提升到了98.1%,满足了起重机工业场景下的损伤检测要求。
In the aspect of crane damage detection,image-based surface corrosion detection is an important emerging technology.Due to the complex structure and shape of hoisting machinery,the crane images obtained by unmanned aerial vehicles and other equipment will contain a large number of shadows and black background objects,which are easily mistaken for rust.Moreover,cranes are high-end equipment,and the sample size of corrosion damage is often insufficient,which aggravates the difficulty of corrosion detection.To solve the above problems,the author proposes a rust detection algorithm for small sample images of cranes based on two-stage network.Specifically,a design of a second-order end-to-end learning network is proposed,which includes segmentation and decision-making.The segmentation network is trained by using the label of corrosion damage,and the decision-making network further judges whether the input picture contains corrosion damage according to the features extracted by the segmentation network.The network is optimized by upgrading the two-stage network from a structure similar to VGGNet to a residual structure similar to ResNet.The experimental results show that the proposed method can effectively improve the accuracy of image corrosion detection,and the average accuracy of the model is increased from 94.2%to 98.1%on the self-built small sample image data set of cranes,which meets the requirements of damage detection in crane industrial application.
作者
王华
张燕超
吴波
翟象平
魏明强
Wang Hua;Zhang Yanchao;Wu Bo;Zhai Xiangping;Wei Mingqiang
出处
《起重运输机械》
2022年第19期47-55,共9页
Hoisting and Conveying Machinery
关键词
起重机
小样本学习
二阶段网络
语义分割
锈蚀检测
crane
small sample learning
two-stage network
semantic segmentation
rust detection
