基于分组卷积和特征图级联的轻量级目标检测

Light-weight object detection network based on group convolution and feature maps cascade

针对Pelee轻量级目标检测网络中参数量和计算量较多、检测精度较差等缺陷,提出了基于分组卷积和特征图级联的轻量级目标检测网络GCPelee。首先,利用分组卷积替换检测模块中的标准卷积形式以减少模型参数量和计算量;其次,在检测模块上应用特征图级联,将感受野较大的特征图包含的信息传递至感受野较小的特征图,提升后者的感受野大小。实验结果表明,优化后的GCPelee模型参数量和计算量均得到减少,检测精度得到了提升。

To solve the shortcomings such as a large number of parameters,a large number of FlOPs and poor detection accuracy in the Pelee light-weight object detection network,this paper proposed an improved version named GCPelee based on grouped convolution and feature maps cascade.Firstly,it reduced the amount of model parameters and FLOPs by replacing normal convolution in the detection module with group convolution.Secondly,it applied feature maps cascade on the detection module to transmitted the information contained in the feature maps with a large receptive field to the feature maps with a small one,which would enlarge the receptive field of the latter.The experimental results show that the GCPelee model gets higher detection accuracy with less parameters and less FLOP.