基于改进YOLOv3的轻量化神经网络算法研究

对于小样本数据集，YOLOv3神经网络框架在训练时存在特征利用率和特征传递效率低的问题，其网络性能得不到充分利用，为解决这些问题，该研究提出基于改进的YOLOv3轻量化神经网络模型，该网络模型将YOLOv3基础框架中的ResNet残差网络结构改为DenseNet的密集串联结构，并将多尺度输出结构删减到2个.在自制麻将子数据集上的实验表明，改进YOLOv3的神经网络的每秒计算帧数(FPS)对比改进前提升了119.03%，预测目标与实际对象交并比(IoU)在0.5以上的平均检测精确度(mAP-50)提升了2.45%.将改进模型推广至开源数据集Kaggle以及Caltech上，改进模型相比原模型的每秒计算帧数分别提升了124.39%、140.05%，预测目标与实际对象交并比在0.5以上的平均检测精度分别提升了12.5%、5.34%.

Research on lightweight neural network algorithm based on improved YOLOv3

For the small sample datasets， the YOLOv3 neural network framework has problems of low feature utilization and low feature transfer efficiency during training， so its network performance is not fully utilized. To solve these problems， this paper proposed an improved YOLOv3 lightweight neural network，which changes the ResNet residual network structure in the YOLOv3 infrastructure to DenseNet's dense tandem structure and reduces the multi-scale output structure to two. Experiments showed that the FPS of the improved YOLOv3 neural network increased by 119.03% and the mAP-50 increased by 2.45%on the homemade mahjong dataset. Extendeding the improved model to the open source datasets like Kaggle and Caltech， the FPS of the improved model increased by 124.39% and 140.05% respectivelywhen compared with the original model， and mAP-50 increased by 12.5% and 5.34%respectively.