Ship Detection And Classification Based On Deep Convolutional Neural Network For High Resolution Remote Sensing Image

The detection of ships is important for civil field,commercial field,military and other fields.Ship detection can not only make important contributions to the regulation of the marine sector,but also affect the economic and territorial security of the country.For example,the relevant departments can maintain maritime traffic safety by monitoring specific sea areas and specific ports,the military can quickly acquire the military deployment and strength of the enemy by detecting certain important military ports.Therefore,it is very important to study how to quickly and accurately detect ship targets.With the rapid development of remote sensing technology,especially the successful launch of high-series satellites,the spatial resolution of remote sensing images is higher and higher,so we can use high-resolution remote sensing images to perform ship detection tasks,but at present,most of the researches on ship detection using high-resolution remote sensing images adopt traditional feature extraction and machine learning methods,or a combination of traditional sliding window and shallow convolutional neural network.The traditional method can not make full use of the rich spectral information and spatial information provided by high-resolution remote sensing images,so the target detection accuracy is not high.Some methods use traditional sliding window method to find candidate regions and shallow convolutional neural network to distinguish candidate regions.However,these methods use a traditional sliding window method to generate a large number of candidate regions,so the detection speed is relatively slow,and the shallow convolutional neural network can not fully extract some important features of the ships,so the accuracy of ship detection of remote sensing images is not high.Based on the above,this paper proposes a HS-CNN model based on deep convolutional neural network to realize high-resolution visible light remote sensing image ship detection task for high-resolution visible light remote sensing images from Google Earth.For the first time,the model combines an SVM classifier with a deep convolutional neural network for ship detection.First,the selective search algorithm is used to find all regions of interest in the image,then HOG+SVM is used to exclude some non-ship regions,and finally a 16-layer deep convolutional neural network is used for final discrimination.The experimental results show that the selective search algorithm used in the high-resolution remote sensing image ship detection model proposed in this paper reduces the amount of computation caused by the traditional sliding window,and then uses HOG+SVM to reduces the further according to the specific probability,and the deep convolutional neural network model can extract the characteristics of the remote sensing image ship more accurately than the traditional machine learning methods and the shallow neural network model to make the final judgment,achieve faster and more accurate results in the ship detection task.Moreover,in view of the fact that there are few studies on ship classification of high resolution visible remote sensing images and the classification accuracy is not high.This paper will use the deep convolutional neural network to complete the high-resolution remote sensing image ship classification task.However,there is currently no open source data for ship classification of high-resolution visible light remote sensing images that can be directly used on the network,so there is not enough data for ship classification tasks.To this end,this paper proposes a high-resolution visible light remote sensing image ship classification method based on transfer learning,and based on Google Earth's high-resolution remote sensing image to establish a simple ship classification data set,which can make good use of the powerful memory,learning ability and feature space expression ability brought by deep convolutional neural network,and fully extract the characteristics of various ships.Finally,this paper uses a variety of source domain datasets to complete the pre-training of deep convolution neural network,and uses ship data sets to fine-tune the network.The effectiveness of the proposed method is verified by comparing the classification results with those without transfer learning.