改进DeepLabV3+网络的指针轨迹图像识别
Pointer Trajectory Recognition with Improved DeepLabV3+Network
袁帅 1蒋强 1饶兵2
作者信息
- 1. 沈阳理工大学 自动化与电气工程学院,沈阳 110159
- 2. 沈阳天眼智云智能技术研究院有限公司,沈阳 110179
- 折叠
摘要
指针式机械记录仪通常用于记录精密设备运输过程中的震动轨迹图像,为了更好地监测运输过程中车辆颠簸对仪器设备的影响,提出一种改进DeepLabV3+网络的指针轨迹图像语义分割方法.首先将骨干网络替换为MobileNetV3,实现模型的轻量化;然后将解码器中 4倍上采样替换为2 次2 倍上采样,增强图像中像素的连续性,使预测结果更接近原始图像.在自制数据集上进行对比实验,结果表明:改进DeepLabV3+网络的平均交并比(MIoU)达到85.84%,比原始DeepLabV3+网络提高了 3.57%,单位时间内检测图片数量(FPS)提高了3.58 s-1;改进DeepLabV3+网络在识别精度和速度上具有明显的优势,可为精密仪器检测提供数据支持.
Abstract
Pointer-type mechanical recorders are usually used to record images of vibration trajecto-ries during the transportation of precision equipment.In order to better monitor the impact of vehi-cle bumps on instruments and equipment during transportation,a semantic segmentation method of pointer trajectory images with improved DeepLabV3+network is proposed.Firstly,the backbone network is replaced with MobileNetV3 to realize the lightweight of the model.Then the 4-fold up-sampling in the decoder is replaced with 2 times 2-fold upsampling to enhance the continuity of pix-els in the image,which makes the predicted results closer to the original image.The results show that the average intersection ratio(MIoU)of the improved DeepLabV3+network reaches 85.84%,which is 3.57%higher than that of the original DeepLabV3+network,and the number of detected images per unit time(FPS)increases by 3.58 s-1.The improved DeepLabV3+network has obvi-ous advantages in recognition accuracy and speed,which can provide data support for precision in-strument detection.
关键词
改进DeepLabV3+/语义分割/轨迹图像识别/轻量化Key words
improved DeepLabV3+/semantic segmentation/trajectory image recognition/light-weight引用本文复制引用
基金项目
辽宁省教育厅科学研究经费项目(LG202014)
出版年
2024
NETL
NSTL
万方数据