摘要
视觉传感图像目标的特征变化幅度普遍较大,类别间样本不均衡,为了提高视觉传感图像关键目标识别效果,提出基于多数据集深度学习的视觉传感图像目标增强识别方法。采用颜色特征提取识别目标的特征量,通过计算视觉传感图像目标与背景区域的纹理特征差异,构建具有多特征参数的卷积神经网络模型。采用指数Laplace损失函数降低模型中类内特征变化幅度,调整不同类中心的特征间距离。结合自适应分块标记完成目标增强识别全过程。以火焰和车辆监控视觉传感图像为分析对象,设计测试实验。结果表明:所提方法的视觉传感图像增强效果较好,能够有效识别出视觉传感图像中火灾发生的异常行为和车辆的驾驶行为,且目标识别耗时低于6.2 s,证明所提方法的实际应用价值较高。
The feature variation of visual sensing image objects is generally large,and the samples among categories are unbalanced.In or-der to improve the recognition effect of key objects in visual sensing images,a method of target enhancement recognition in visual sensing images based on multi dataset depth learning is proposed.Color feature is used to extract the feature quantity of object recognition,and the convolution neural network model with multiple feature parameters is constructed by calculating the texture feature difference between the object and the background area of the visual sensing image.Exponential Laplace loss function is used to reduce the variation range of intra class features in the model and adjust the distance between features of different centers.The whole process of target enhancement recogni-tion is completed by combining adaptive block marking.Taking the visual sensing images of flame and vehicle monitoring as the analysis object,the test experiment is designed.The results show that the visual sensing image enhancement effect of the proposed method is good,the abnormal behavior of fire and vehicle driving behavior in the visual sensing image can be effectively recognized,and the target recogni-tion time is less than 6.2 seconds,which proves that the proposed method has high practical application value.
作者
董涛
杨宝华
DONG Tao;YANG Baohua(School of Information Engineering,Liaodong University,Dandong Liaoning 118000,China)
出处
《传感技术学报》
CAS
CSCD
北大核心
2024年第1期64-70,共7页
Chinese Journal of Sensors and Actuators
基金
2022年辽宁省自然科学基金项目(2022-KF-12-13)。
