Research On Detection Of Ships In Remote Sensing Images Based On Deep Learning

With the advantages of large transportation volume and low transportation cost,ships have become an important means of transportation at sea,and play an important role in international trade.Extracting ships based on high-resolution remote sensing images can not only ensure the safe navigation of ships,but also improve the navigation efficiency of waterways.Deep learning algorithm can process massive remote sensing images,which makes it possible to intelligently detect ships in remote sensing images.Therefore,this paper studies ship object detection based on deep learning in remote sensing images.(1)Aiming at the problem that the complex and huge object detection model has poor availability on the platform with limited computing resources and the detection results of the existing object detection models lack ship direction information,a lightweight convolutional neural network model of ship object detection and direction recognition model is proposed.Specifically,although the complex and huge model has high detection accuracy,the model complexity is high and the computing resource requirements are relatively high,which limits the use of the model.At the same time,due to the unique imaging perspective of remote sensing technology,remote sensing images not only contain ship category and position information,but also rich direction information of ship objects,while existing model detection results lack ship direction information.In view of the above problems,the model constructed in this paper is based on the Faster-RCNN algorithm.First,one-shot feature aggregation modules and depthwise separable convolutions are used to form an efficient and lightweight feature extraction network;At the same time,the K-Means++ algorithm is introduced into the RPN network to form K-RPN to generate high-quality region proposals.Without introducing additional parameters,this paper transforms the problem of direction recognition of ships in remote sensing images into a classification problem and completes the estimation of the four directions of east,south,west and north while completing the detection task.(2)Aiming at the high cost of object level labeling,a weakly supervised learning ship object detection model is proposed to reduce the cost of data labeling.The premise of strong supervised object detection model to obtain good performance is that a large amount of data needs to be labeled in the form of object level annotation to complete the training of the model.In the object level labeling process,not only the category information of the object,but also the position information of the object needs to be labeled,which requires a lot of labeling time and cost.Image level annotation only needs to label the category information of objects,which can save a lot of annotation costs.To solve this problem,a weakly supervised learning ships detection model is proposed.Based on the idea of weakly supervised learning,combined with the principle of convolutional neural network and class activation mapping,two weakly supervised learning methods VGG16-CAM and Res Net50-CAM are constructed by using VGG16 and Res Net50 convolutional neural networks respectively.(3)Taking cargo ships in remote sensing images as research objects,the characteristics of three typical cargo ships are sorted out: bulk carriers with unique hatches,container ships filled with containers,tankers with tubing.Then,based on Google Earth Image,a dataset of ship object detection and direction recognition from remote sensing image is constructed,which contains the category,location and direction information of cargo ship as well as the direction information of east,south,west and north.A ship object detection dataset with remote sensing image containing cargo ship category information is constructed by image level annotation.(4)Based on the self-built datasets,the proposed models are trained and tested,and the experimental results are evaluated.The experimental results show that the lightweight convolutional neural network model of ship object detection and direction recognition is more suitable for the platform with limited computing resources in terms of model size(110MB),detection accuracy(91.96%),and detection speed(reasoning time of a single remote sensing image is 46 ms).At the same time,the model is also valid for single source high-resolution remote sensing images except Google Earth Image.Weakly supervised learning ship object detection model of VGG16-CAM model and Res Net50-CAM model detection accuracy m AP are 72.06% and 80.56%respectively,among which,Res Net50-CAM model has the highest detection accuracy,which is more suitable for weakly supervised model training and used for ship object detection in remote sensing images.The lightweight convolutional neural network model of ship object detection and direction recognition model constructed in this paper further excavates the semantic information of ships in remote sensing images,and reduces the requirements for computing resources;the weakly supervised learning ship object detection model constructed reduces the data label cost.This work is expected to promote the improvement of the level of intelligent interpretation of remote sensing images,and has broad application prospects.