基于目标检测的VOC泄漏区域识别技术研究  

Research on VOC leakage area identification based on target detection

作　　者：杨承霖 陈海永 岳学智 李华曜[2] 邓立齐 郭东歌 王海超 刘欢[2] YANG Chenglin;CHEN Haiyong;YUE Xuezhi;LI Huayao;DENG Liqi;GUO Dongge;WANG Haichao;LIU Huan(Hanwei Electronics Group Corporation,Zhengzhou 450001,China;Optics Valley Laboratory,School of Integrated Circuits,Huazhong University of Science and Technology,Wuhan 430074,China)

机构地区：[1]汉威科技集团股份有限公司,郑州450001 [2]华中科技大学集成电路学院湖北光谷实验室,武汉430074

出　　处：《激光技术》2024年第6期922-930,共9页Laser Technology

基　　金：华中科技大学湖北光谷实验室创新科研项目(OVL2023ZD003);郑州市重大科技专项基金资助项目(2021KJZX0061)。

摘　　要：为了解决红外气体成像仪在挥发性有机化合物(VOC)泄漏识别中存在的误识别率高、漏检率高、算法执行效率低以及模型泛化能力差等问题,提出了一种基于运动特征增强的VOC泄漏区识别方法。采用视频序列投影变化率统计的方法确定视频稳定性判定阈值,提取稳定状态下运动背景和运动前景;采用优化线性拉伸的方法对运动前景进行特征增强和异常值过滤;将运动前景与原始帧进行图像融合,并利用目标检测算法进行VOC泄漏区域识别;通过模型预训练和迁移学习的方法,以烟雾数据集和少量VOC泄漏数据集进行了识别模型训练,并将模型迁移至RK3588S嵌入式开发板上进行了执行效率测试。结果表明,该算法在交并比为0.5的情况下,平均精度均值为0.88;交并比在0.5~0.95范围内,平均精度均值为0.51,单帧平均识别时间为49 ms,具有较高的识别精度和识别效率,能够满足实时监测需求。本文中的算法能够保持稳定的模型性能且具有一定的抗干扰能力,为VOC泄漏识别提供了一定的参考。To solve the problem of high misidentification rate,high missed detection rate,low algorithm execution efficiency,and poor model generalization ability in the recognition of volatile organic compound(VOC)leakage area of infrared gas imager,a VOC leakage area recognition method based on motion feature enhancement was proposed.The video stability threshold was determined by using the statistical method of projection change rate of video sequence,and the moving background and moving foreground were extracted under stable state.Optimized linear stretching was used to perform feature enhancement and outlier filtering on the moving foreground.The motion foreground was fused with the original frame,and VOC leakage area identification was performed using the target detection algorithm.Through the method of model pre-training and transfer learning,the smoke dataset and a small amount of VOC leakage dataset were used to train the recognition model,and the model was transferred to the RK3588S embedded development board for execution efficiency test.Experimental results show that the mean average precision of the proposed algorithm is 0.88 when the intersection over union ratio is 0.5,and the mean average precision is 0.51 when the intersection over union ratio ranges from 0.5 to 0.95.The average recognition time of a single frame is 49 ms,which has high recognition accuracy and recognition efficiency,and can meet the requirements of real-time monitoring.The algorithm in this article can maintain stable model performance and has certain anti-interference capabilities providing some reference for VOC leak identification.

关 键 词：图像处理 泄漏区域识别 目标检测 挥发性有机化合物泄漏 红外气体成像仪 

分 类 号：TN219[电子电信—物理电子学]