摘要
为实现SSD算法模型对人脸的目标检测,采用公开人脸数据集FDDB对网络模型进行重新训练改进。通过训练时输入不同的人脸数据集来优化网络训练结果。针对人脸检测训练过程中的过拟合问题,通过降噪自编码器的方法,在输入数据集中加入负样本,在训练模型中生成噪声。通过L1正则化产出稀疏模型,稀疏模型具有更好的特性去处理高维的数据特征以增强模型的泛化能力,实现在网络迭代训练过程中降噪的效果,防止模型陷入过拟合。然后通过非极大值抑制算法(NMS)使候选框确定为最终的人脸检测窗口进行人脸检测。在训练平台MXnet下的实验结果表明,加入噪声后的人脸检测模型的mAp(mean average precision)性能提高至0.997,同时在提高遮挡、光照、小目标等检测的鲁棒性的情况下,仍保持较快的收敛速度。
In order to realize the face target detection based on SSD(single shot multibox detector)algorithm,public face data set FDDB is used to retrain and improve the network model.The network training results are optimized by inputting different face data sets during training.For solving the problem of over-fitting in the training process of face detection,negative samples are added into the input data set by the method of self-encoder to generate noise in the training model.The sparse model is produced by L1 regularization,which has better characteristics to deal with high-dimensional data features,so as to enhance the generalization of the model,achieve the effect of noise reduction during network iteration training,and prevent the model from falling into over-fitting.Then the candidate box is determined as the final face detection window for face detection by non-maximum suppression(NMS).The experiment on the training platform MXnet shows that the mAp(mean average precision)of the face detection model with noise is improved to 0.997.At the same time,the robustness of occlusion,illumination and small target detection is improved,and the convergence speed is still fast.
作者
杨璐
吴陈
YANG Lu;WU Chen(School of Computer Science&Technology,Jiangsu University of Science and Technology,Zhenjiang 212003,China)
出处
《计算机技术与发展》
2019年第10期181-185,共5页
Computer Technology and Development
基金
国家自然科学基金(61572242)
关键词
卷积神经网络
SSD算法
NMS算法
正则化
人脸检测
convolution neural network
SSD algorithm
NMS algorithm
regularization
face detection