YOLOv3作为开源的目标检测网络与同时期目标检测网络相比,在速度和精度上有着明显的优势。由于YOLOv3采用了新型的全卷积网络(FCN)、特征金字塔网络(FPN)和残差网络(ResNet),因此对硬件配置要求较高,导致开发成本过高,不利于工业上的应用普及。在嵌入式平台上普遍使用YOLOv3tiny进行检测,虽然计算量较小,但是检测效果远不如YOLOv3。为了解决在嵌入式平台上YOLOv3检测速度低的问题,提出一种基于YOLOv3的简化版网络,与YOLOv3不同的是,在保留了对特征提取有较大帮助的FCN、FPN以及ResNet的同时,尽可能减少每层的参数量和残差层数,并尝试加入了密集连接网络空间金字塔池化。实验结果表明,该网络的参数量和检测速度大幅优于YOLOv3,且平均精度比YOLOv3tiny在PASCAL VOC2007、2012数据集上有明显的提升。

As an open source object detection network, YOLOv3 has obvious advantages in speed and accuracy compared with the object detection network of the same period. Because YOLOv3 adopts a new type of full convolutional network (FCN), feature pyramid network (FPN), and residual network (ResNet), it requires high hardware configuration, leading to high development cost, which is not conducive to the popularization of industrial applications. Therefore, YOLOv3tiny is generally used for detection on embedded platforms. Although the calculation amount is small, the detection effect is far less than YOLOv3. To solve the problem of low detection speed of YOLOv3 on embedded platforms, a simplified version of the network based on YOLOv3 is proposed. Unlike YOLOv3, FCN, FPN, and ResNet, which are helpful for feature extraction, are retained as much as possible. the number of parameters and residual years of each layer is recued, and attempts are made to join densely connected networks and spatial pyramid pooling. Experimental results show that the number of parameters and detection speed of this network is much better than YOLOv3, and the mean average precision is a significant improvement compared to YOLOv3tiny in terms of in the PASCAL VOC2007 and 2012 datasets.