Research On Image Based Ship Detection And Classifcation

With the development of intelligent technology,ship image target detection and recognition technology has gradually become a research hotspot,and ship image has some special features.Firstly,the working environment of ship target is more complicated.For example,in the nearshore background,the ship target is often mixed with various interfering objects.Secondly,the sea surface environment is relatively open,and the size variation range of ship target is relatively large.These problems bring severe challenges to the methods of target detection and classification.This paper summarizes the main methods and analyzes their shortcomings by referring to the literature related to ship target detection and recognition,and proposes a method of ship image target detection and recognition based on Mixed Attention Model.The main work of this paper is as follows:This paper introduces the basic theory of target detection and recognition,including various network layer structures and gradient descent optimization algorithm,which provides a theoretical basis for the network design and optimization algorithm in the following chapters;A ship target detection and recognition network based on training optimization YOLOV3 is proposed.The hybrid training method of transfer learning and non-transfer learning further trains the image low-level feature extractor with a lower learning rate,making it more suitable to extract the low-level features of ship image.The method of learning rate attenuation ensures that the network parameters update quickly in the early stage of training,and makes the network converge more easily in the later stage.Early Stopping real-time calculation of training set and verification set loss,to avoid overfitting problems.This paper analyzes the sample imbalance problem of ship image and proposes a network of ship object detection and classification based on Pixel Attention Model.The Pixel Attention Model is a generated antagonistic network structure,including generator and discriminator.The model generates a mask map for the ship image input and attenuates the pixel value of the non-target ship area to enhance the sensitivity of the network to the target area.Considering pixel sample imbalance problem and feature sample imbalance problem in ship image,a ship target detection and classification network based on Mixed Attention Model is proposed.On the basis of the Pixel Attention Model,the Mixed Attention Model adds the Feature Attention Model to attenuate the useless feature and the useless convolution kernel by calculating the feature weight in the two directions of breadth and depth,so as to enhance the overall performance of the network.Summarizes the main works,analyses the application prospects and existing problems.