基于深度学习的行人属性辨识研究.pdf

第**卷第**期 机 械 工 程 学 报 Vol.** No.* 2 0 * * 年 * 月 JOURNAL OF MECHANICAL ENGINEERING *** 20** DOI ：10.3901/JME.20**.**.*** 基于深度学习的行人属性辨识研究 刘嘉琳 (哈尔滨工业大学（威海） 山东) 摘要：本文提供了一个采用更多消耗函数方法的网络模式,将交叉熵损耗函数和经过修改的监控信息中心损失函数方法结合, 对模型参数加以了修改。最后,所提取出的特点将更为明显,使同一个人的特点也更为紧凑。与只用交叉熵损失函数培训的网 络比较下,对交叉年龄数据集的检测结果做出了改善,使采用 WEB 技术开发的行人属性辨识应用领域也能够正常进行。与其 他类型的应用程序比较,该应用领域具备了跨平台、轻量化的特性,其功能主体涉及图像采集、行人属性辨识以及数据存储模 块等,可广泛应用于日常生活辨识任务。 关键词：深度学习；行人属性识别；多损失函数 Research on pedestrian attribute recognition based on deep learning LIU Jialin (Harbin Institute of Technology (Weihai), Shandong) Abstract ：This paper provides a network model using more consumption function method, which combines the cross entropy loss function with the modified monitoring information center loss function method to modify the model parameters. Finally, the extracted features will be more obvious, making the characteristics of the same person more compact. Compared with the network trained only by cross entropy loss function, the detection results of cross age data set are improved, so that the application field of pedestrian attribute identification developed by web technology can also be carried out normally. Compared with other types of applications, this application field has the characteristics of cross platform and lightweight. Its functional main body involves image acquisition, pedestrian attribute recognition and data storage module, which can be widely used in daily life recognition tasks. Key words ：Deep learning; Pedestrian attribute identification; Multiple loss function 位置，实施刑事逮捕。并将行人属性识别监控系统 * 1