Research On Lightweight Real-time Target Detection Depth Network Based On Attention Mechanism

With the rapid development of convolutional neural network,computer vision technology has made a great breakthrough.In the field of visual inspection,there have been many classic high-precision models such as YOLO and Fast-RNN.However,these network models also have two problems: one is that the parameters and floatingpoint calculations of the classical target detection model are too large to be deployed in mobile devices with weak performance;the other is that the performance of the network model declines linearly with the reduction of network model parameters and floatingpoint calculations,that is,the detection accuracy is insufficient after the network is lightweight.These problems limit the application and detection effect of target detection algorithm.To solve the above problems,this paper improves the mainstream target detection network YOLOv3,and proposes a lightweight target detection network YOLO-SNettiny.The network model mainly carries out research and innovation in two aspects:network model lightweight and network model detection accuracy improvement.For network model lightweight: first,the backbone network of YOLOv3 model is directly replaced by the network structure of Shufflenet v2 lightweight feature extraction.The second is to replace the standard convolution commonly used in neck network and detection head network with lightweight depth convolution.These two methods greatly reduce the amount of parameters and floating-point calculations in the network model.Aiming at improving the detection accuracy of network model: first,a lightweight neck network structure PAN-tiny is proposed for top-down and bottom-up bidirectional feature fusion.The second is to propose a new hybrid attention mechanism module to enhance the attention of the network to important information.Thirdly,a new lightweight decoupled detection head Lite is proposed to refine the classification and regression tasks.These three methods alleviate the situation that the performance of the target detection model degrades greatly after the model is lightweight.In order to verify the performance of YOLO-SNet tiny,experiments are conducted on the embedded device Jetson nano Bo1 CPU platform with COCO 2017 as the training and verification data set.The experimental results show that YOLO-SNet-tiny achieves 30.4% of map(0.5)on the COCO data set.Compared with the leading-edge model YOLOv4-tiny,under the same resolution,YOLO-SNet-tiny only loses 10%map(0.5)less accuracy than YOLOv4-tiny.The model reasoning speed is about 3 times as fast as that of YOLOv4-tiny.The floating-point operation amount is only about 1/14 of that of YOLOv4-tiny,and the network parameter amount is about 1/6 of that of YOLOv4-tiny.