Lightweight target detection algorithm based on multi-scale feature fusion
In order to solve the problem of many parameters and high computational complexity in the YOLOv5 target detection model,which cannot meet the needs of edge devices for intelligent computation and real-time feedback.A lightweight target detection algorithm based on multi-scale feature fusion is proposed.Firstly,to address the problem of large number of parameters and high computational complexity of the standard convolution module,a phantom convolution-based feature extraction convolution module is proposed to replace the feature extraction module of the original model,which reduces the number of parameters and the computational amount of the model under the premise of maintaining the detection accuracy.Then the ShuffleNetv2_2 downsampling module is designed to further reduce the number of parameters of the algorithm.Secondly,to address the problem of insufficient feature extraction ability after model lightweighting,the low-dimensional features are fully fused into the Neck network and cross-layer cascade is added to reduce the loss of shallow semantics,which enhances the expression of the target features and improves the detection efficiency of the model at the same time.Finally,the LAM attention fusion module is proposed to provide the model's Neck network with a richer semantic feature map.The experimental results show that the improved model has fewer parameters and less computation than the original model,and the detection accuracy is improved by 2.1%and 2.4%in the Pascal VOC and MS COCO datasets,respectively.
NETL
NSTL
万方数据
