Research On Lightweight Convolutional Neural Network And The Application

In recent years,as deep learning software and hardware technologies have become more mature,the artificial intelligence civilian market based on deep learning has steadily developed and has a bright future,especially in the fields of object classification,face recognition,target detection,data analysis,and unmanned operation.With its powerful performance and efficient implementation advantages,artificial intelligence has successfully penetrated into all aspects of social life and has attracted widespread attention.Among them,target detection is one of the important topics of deep learning.And how to use as few resources as possible to implement the target detection function faster and better,and apply it to the scene with the least restrictions,that is,the lightweight transplantation of the target detection task,is a topic worthy of academic research.This article is based on the realization of autonomous positioning of drones in a GPS-free environment.It requires the use of vision sensors for data acquisition,the use of airborne embedded devices to run a lightweight target detection network to process visual information to obtain the relative position of the target object,and then to achieve autonomous positioning of the drone.The main research contents of the paper are as follows:First,this article discusses the development of classic convolutional neural networks and target detection networks,studies commonly used lightweight strategies and classic lightweight networks,studies the feasibility of lightweight transformation of target detection networks,and finally chooses to improve lightweight YOLOv3.Subsequently,this article studies YOLOv3 in depth,uses new strategies to improve YOLOv3's backbone network,and combines existing Dense Net and SPP-Net theories to improve YOLOv3's feature fusion network,and then adjusts the weight of YOLOv3's loss function to comprehensively propose a new target detection network LDC-YOLO,on the PASCAL VOC dataset,proved the advantages of the network in terms of detection speed and weight size.Finally,this paper establishes a visual positioning coordinate system,analyzes the camera projection imaging principle,uses a checkerboard to calibrate and solve the camera's internal parameters and distortion parameters.In real scenes and simulation environments,the target object information obtained by the LDC-YOLO network is used to perform The coordinate system is transformed and the relative position information of the drone is obtained.The experimental detection of the autonomous positioning function algorithm shows that the algorithm is effective and can meet the needs of UAV autonomous positioning.