| In recent years,with the increasing demand for prediction and classification accuracy,corresponding deep neural network models with deeper and more complex structures have been gradually proposed and applied.At the same time,the improvement of the network accuracy performance also brings the increase of the size and operation amount of the neural network model,and the decrease of the operating efficiency.With the widespread popularity of mobile and embedded devices and the increase in edge computing scenarios,in practical applications,deep neural network models are often required to ensure both high accuracy and good real-time performance.Current embedded edge computing devices are difficult to meet actual needs in terms of computing power and storage resources,so only good compromises and compromises are made in terms of accuracy,network parameter size and speed,that is,limited computing power and resources The optimal model accuracy performance is achieved in order to better apply the deep neural network to the edge calculation of the mobile terminal.This article focuses on the lightweight improvement of the target detection model YOLOv3.By introducing the idea of the Mobile Net model and other targeted measures,the improved YOLOv3 model can better achieve real-time and efficient target detection on embedded terminal devices,thereby serving as an edge Applications in computing scenarios provide possible solutions.The main work includes:1.Combining Mobile Net and YOLOv3,a lightweight neural network structure Mobile Net＿YOLOv3 is proposed,so that the proposed model can be deployed to run on an embedded mobile terminal platform in real time,and the average detection accuracy is improved compared with other lightweight models.The Mobile Net model is a lightweight deep learning separation model developed by Google.It is suitable for mobile embedded environments and has low latency while ensuring that the accuracy can meet the requirements.YOLOv3 is the latest improved version of the YOLO model.The model combines the YOLOv2 and Faster R-CNN anchor mechanism theory,and has the advantages of these two algorithms.It inherits the YOLO algorithm in terms of running speed and can be like Faster.R-CNN is just as accurate.However,for many embedded mobile platforms,the YOLOv3 model has a large amount of computation.Therefore,this paper uses Mobile Net's ideas and deep separable convolution structure to improve the YOLOv3 model,and proposes the Mobile Net＿YOLOv3 network model.The proposed new model makes full use of the advantages of these two network models,and achieves a better compromise in terms of model recognition accuracy and operating efficiency.The test results on the international general dataset show that the proposed Mobile Net＿YOLOv3 model is comparable to the original YOLOv3 model in performance but has a significant reduction in computing capacity.Compared with other lightweight models,the performance and computing efficiency are improved.2.An improved YOLOv3 model for small targets and occlusion is proposed.Although Mobile Net＿YOLOv3 has improved in real time compared to the original YOLOv3 model,there are still limitations in the detection accuracy of small objects and occlusion objects.To solve this problem,this paper designs a multi-scale feature fusion attention network(MSFAN),which can effectively reduce the loss of feature information after image convolution and improve the detection accuracy.At the same time,it proposes data oversampling and enhancement strategies for the problem of small samples in the dataset,which further improves the model's ability to detect small targets.The experimental results show that the average detection accuracy of the MSFAN model is 5.2% higher than that of the original model YOLOv3 on the PASCAL VOC dataset,and the accuracy of the two is equivalent on the MS COCO dataset,and the speed is nearly doubled. |
