Research On Ship Identification Method Based On Depth Convolution Neural Network

With the continuous development of target recognition technology in the field of computer vision and the rise of deep convolutional neural networks,the application of deep learning in target recognition gradually exceeds the traditional target recognition methods.Some of these emerging neural network-based algorithms have high accuracy,some have fast recognition speed,and others are compatible.How to better apply these methods to the actual needs requires continuous research and exploration,which has become a research hotspot in recent years.This article takes ships as research objects,and conducts research on different target recognition algorithms used in ship recognition in different scenarios.In order to accurately identify ship targets in remote sensing images,a target recognition algorithm Dense-YOLOv3 is proposed.In addition,in order to accurately recognize the ship type in the visual image under the condition of real-time detection speed,a lightweight target recognition algorithm was proposed by improving the YOLOv3 algorithm.In view of the above two situations,the specific research content of the paper is as follows:Firstly,This paper summarizes the development status of target recognition algorithms based on deep convolutional neural network in recent years,and then briefly introduces several classical neural network target recognition algorithms.After comparing the advantages and disadvantages of classical algorithms,The YOLOv3 target recognition algorithm is selected as the basis of this paper.According to different regularization methods and optimization methods in the neural network,the neural network is established for experimental comparison,thus determining the optimization method used in the construction of the neural network in this paper.Secondly,for the target recognition scenes with small ship targets,complex backgrounds,and blurred targets in the remote sensing image,the Dense Net network structure that can make full use of the upper network information is added to the selected algorithm YOLOv3,and the Dense-YOLOv3 algorithm is proposed.In addition,for the candidate frames that need to be selected,the K-means ++ method is used to quickly cluster the size of the candidate frames that need to be detected.After that,the performances of Dense-YOLOv3 algorithm,YOLOv3 algorithm,Faster R-CNN and SSD were compared through simulation experiments.Finally,for the target recognition scene with large ship targets and clear images in visual images,a lightweight target recognition algorithm capable of satisfying real-time detection is proposed.Based on YOLOv3,combined with the advantages of deep separable convolution in Mobile-Net lightweight network,four lightweight target recognition algorithms are designed,and their advantages and disadvantages are compared experimentally,and two kinds of comprehensive performances are selected.The better algorithm is compared with the detection effect of YOLOv2-tiny and YOLOv3-tiny in ship images,which verifies the feasibility of the lightweight target recognition algorithm designed in this paper.