摘要
针对审计调查中人脸图像识别应用场景,首先模拟了一个相应环境干扰的人脸数据集,增加数据的多样性;在预处理阶段,研究采用自适应直方图光照平衡和模拟眼镜遮盖的改进方法进行了图像增强处理;在模型训练阶段,提出一种最优权重再重载的模型训练算法。对比实验结果表明,经过图像增强预处理的改进和最优权重再重载的模型训练,提高了应用场景中人脸识别的准确率、鲁棒性和效率。
According to the application scenarios of face recognition in the audit investigation, a face data set with corresponding environmental interference was simulated in advance to increase the diversity of data. In the preprocessing stage, transformation operations such as adaptive histogram light balance, glasses concealment, and horizontal flipping were done, which improved the accuracy and robustness of the training model. In the pre-processing comparison experiment, the accuracy of the model increased by5.7% after a series of pre-processing. In the model training stage, a model training algorithm with optimal weights and reload is proposed. Experimental results show that this training algorithm has significant advantages over traditional training algorithms in termsof the convergence speed and final accuracy of model training.
作者
王海荣
李伟波
万权
鄢华
向锐
WANG Hai-rong;LI Wei-bo;WAN Quan;YAN Hua;XIANG Rui(Hubei Key Laboratory of Intelligent Robot,Wuhan 430205,China;School of Computer Science&Engineering,Wuhan Institute of Technology,Wuhan 430205,China)
出处
《电脑知识与技术》
2021年第10期12-18,共7页
Computer Knowledge and Technology
基金
湖北省高校产学研合作重点资助项目(C2010033)。
关键词
深度学习
人脸识别
审计调查
图像增强
模型训练
deep learning
face recognition
audit investigation
image enhancement
model training