基于深度学习的遮挡人脸快速检测识别算法

Fast Detection and Recognition Algorithm for Occluded Faces Based on Deep Learning

针对机场复杂环境,基于深度学习改进了机场人脸在目标遮挡和多尺度目标情况下的快速检测识别算法。首先,针对人脸遮挡问题,提出了使用区域交叉熵损失函数的网络对人脸图像进行分类,直接将池化的特征图的通道数与最终类别数对应;然后,为获得的特征图每一个通道训练分类器,增加模型对于局部区域的敏感程度;最后,针对训练模型层数较深、权重参数太多的问题,引入了奇异值压缩算法,极大减少了模型训练时的计算复杂度,提高了模型检测目标人脸的速度。经过长时间试验验证,机场人脸快速检测识别系统工作稳定,具有较好性能。

Aimed at the airport's complex environment and based on deep learning,the fast detection and recognition algorithm on the airport faces in the case of the occluded target and the multi-scale tar⁃get is improved.Firstly,aimed at the face occlusion problem,a network with regional cross entropy loss function is proposed to classify face images.The channel number of the pooled feature map is di⁃rectly corresponded to the number of final categories.Then,a classifier for each channel of the gener⁃ated feature map is trained to increase the sensitivity of the model to the local area.Finally,aimed at the problem of too deep layers of the training model and too many weight parameters,a singular value compression algorithm is introduced,and the computational complexity during the model training is much reduced.Thus,the model's detection speed of the target faces is enhanced.After the long-term experimental verification,the airport face fast detection and recognition system works stably,and the system has got better performance.