基于卷积神经网络的轻量化目标检测网络

Lightweight Object Detection Network Based on Convolutional Neural Network

针对目前常用的目标检测算法计算复杂度高,在嵌入式平台检测速度低的问题,提出一种适用于嵌入式平台的轻量化目标检测网络(BENet)。首先,该网络在MobileNetv2轻量化网络的基础上加入通道特征交织模块,来设计骨干网络,有效地增强了轻量化骨干网络的特征表达;其次,提出自适应多尺度加权特征融合模块,通过对不同尺度的特征进行权重分配,学习各个尺度特征之间的相关性;最后,尝试引入空间金字塔池化结构来获取不同感受野的上下文信息。在VOC数据集上的实验结果表明:所提BENet在保持较高目标检测精度和检测速度的同时,具有较低的计算复杂度和较小的参数量,更适合应用于嵌入式平台。

Considering the high computational complexity and low detection speed of the common object detection algorithms on an embedded platform,this study proposes a lightweight object detection network(BENet)suitable for embedded platforms.First,the proposed network added a channel feature interweaving module to the MobileNetv2 lightweight network to design the backbone network,which effectively enhanced the feature expression of the lightweight backbone network.Second,an adaptive multiscale weighted feature fusion module was proposed to learn the correlation between the features with various scales by assigning weights to the features with different scales.Finally,we attempted to introduce a spatial pyramid pooling structure to obtain the context information of different receptive fields.The experimental results on the VOC dataset show that the proposed BENet maintains high object detection accuracy and speed while has lower computational complexity and smaller parameters.Additionally,it is more suitable for embedded platforms.