Abstract: Video-oriented deep learning algorithms have high computational complexity and are difficult to meet real-time requirements,which seriously affects their applications in edge computing and real-time systems.Lightweight networks have become one of the research hotspots.Lightweight networks for large networks significantly reduce the scale of the original network parameters and improve the detection speed,but the detection accuracy is had to meet industrial needs.In view of the above problems,this paper proposes an improved lightweight target detection network,which can effectively improve the detection performance while maintaining a small parameter scale.In this paper,the vision transformer(VIT) structure is added to the YOLOv4-tiny backbone network,and the multi-head self-attention mechanism enables the network to extract deeper object features.Using the simplified Bi-FPN,the two detection channels are changed to three detection channels,and the attention mechanism is introduced in the feature map fusion node to improve the model's utilization of image features and the network’s detection accuracy for objects of different sizes.Using Ghost convolution to replace traditional convolution operations,so as to reduce network computational complexity and network parameters.Experimental results on the COCO dataset show that the improved algorithm has significantly improved the detection accuracy of the original YOLOv4-tiny network while keeping the network scale unchanged,it can simultaneously meet the requirements of edge computing and real-time systems for the lightweight and accuracy of deep networks.

Key words: Object detection, Lightweight network, Multi-head self-attention mechanism, Weighted feature fusion