基于YOLOv3的轻量化高精度多目标检测模型

Lightweight High Precision Targets Detection Model Based on YOLOv3

针对当前目标检测模型在边缘设备中的应用占用内存过大、无法达到实时性要求的问题,提出一种基于YOLOv3的轻量化多目标检测模型.采用MobileNet网络进行点卷积和深度可分离卷积运算提取图像特征,显著降低了模型的参数量.同时,为了保证目标检测精度,在训练过程中不仅采用CIOU(completeintersectionoverunion)目标框回归损失函数,而且在损失函数中引入Focal loss,减少正负样本分布不平衡所造成的误差;引入Label Smoothing调整真实样本标签类别在计算损失函数时的权重,有效抑制过拟合问题.经3.5万个实际场景数据训练,本文提出的改进模型在行人和车辆的检测精度上分别达到47.3%和69.67%,模型大小仅为YOLOv3的40%,实现了理想检测精度水平下的模型轻量化.

Aiming at the problem that the current target detection algorithm occupies too much memory in the application of the edge device and cannot meet the real-time requirements,this article proposes an improved lightweight targets detection model based on YOLOv3 algorithm.MobileNet was adopted to carry out point convolution and deep separable convolution for features extracting,which significantly reduces the number of parameters of the model.Meanwhile,in order to ensure the accuracy of target detection,complete intersection over union(CIOU)bounding box regression loss function was used in the training process.In addition,Focal loss was introduced to reduce the errors caused by the unbalanced distribution of positive and negative sample distributions.Moreover,Label Smoothing was taken as an optimized strategy to adjust the weight of the real sample label category in the calculation of the loss function,which is helpful to avoid the overfitting problem.Aftertraining on 35000 actual scene data,the proposed model improves 47.3%and 69.67%detection accuracy of pedestriansand vehicles respectively,and the model size is 40%of YOLOv3,thus achieving the targets detection with lightweight model and satisfying precision.