基于自适应空间特征融合的轻量化目标检测算法

Lightweight Target Detection Algorithm Based on Adaptive Spatial Feature Fusion

针对目前深度学习中单阶段目标检测网络结构复杂、训练困难与在移动与嵌入式设备难以部署的问题,提出了一种基于自适应空间特征融合的轻量化目标检测算法。所提算法以YOLOv4为网络基础框架,采用轻量级MobileNet作为特征提取网络,降低网络深度与训练难度,提高检测速度;采用一种自适应空间特征融合(ASFF)方式改进PANet对多尺度特征融合效果差的不足;通过增加网络的输出维度,利用Gaussian算法对新增维度建模并输出预测框位置的不确定性;最后对位置损失函数进行重新定义,提高位置回归的准确性。所提算法以疫情期间口罩佩戴检测机器人为部署载体,对人脸口罩佩戴情况进行了测试,实验结果表明,所提算法的检测精度达到了95.92%,检测速度达到了19 frame/s,相比于原始算法和其他主流检测算法,更适合部署于移动与嵌入设备实现实时检测。

Aiming at the problems of complex network structure, difficult training and difficult deployment in mobile,and embedded devices of single-stage target detection in deep learning, a lightweight target detection algorithm based on adaptive spatial feature fusion is proposed. The proposed algorithm takes YOLOv4 as the basic framework of the network and uses lightweight MobileNet as the feature extraction network to reduce the network depth and training difficulty and improve the detection speed;an adaptive spatial feature fusion(ASFF) method is used to improve the poor effect of PANet on multi-scale feature fusion;by adding the output dimension of the network, the Gaussian algorithm is used to model the new dimension and output the uncertainty of the position of the prediction box;finally,the position loss function is redefined to improve the accuracy of position regression. The proposed algorithm takes the mask wearing detection robot during the epidemic as the deployment carrier to test the face mask wearing. The experimental results show that the detection accuracy of the proposed algorithm reaches 95. 92% and the detection speed reaches 19 frame/s. Compared with the original algorithm and other mainstream detection algorithms, the proposed algorithm is more suitable for deployment in mobile and embedded devices to realize real-time detection.