Ship Identification Under Complex Sea Conditions Based On Deep Learning

Marine targets have always been an important direction in target recognition,and ships as an important transport carrier and military target at sea have practical significance for automatic recognition of targets at sea.However,in practice,remote sensing satellites often suffer from complex sea conditions,such as cloud occlusion,land background interference.At the same time,because of the high resolution of satellite images,the volume of ship targets is small,these factors will increase the complexity of ship recognition.Based on the research of remote sensing satellite images characteristics,this paper studies the ship recognition method in complex sea conditions based on the Faster R-CNN target recognition method in deep learning.The main research contents are as follows:This paper first describes the current research status of remote sensing satellite image target recognition and deep learning at home and abroad.On this basis,we analyze the problems of ship recognition under complex sea conditions based on deep learning,such as cloud occlusion and land background interference.And the small target size of the ship,these factors will lead to the problem of low recognition accuracy of the ship.Then a ship recognition network was built based on Faster R-CNN,but this method can only recognize ships in the conventional sea conditions and cannot solve the complex sea level interference.To solve the above problems,this paper adopts a multi-level cascade online hard example mining ship recognition network model,which can be divided into four parts: multi-scale training,feature extraction,generate target proposal area and ship classification.Firstly,aiming at the problem of high miss detection rate of small target,multi-scale training strategy is adopted to train multi-scale ship samples into the network.A large number of small target ship samples are added to the training samples,so that the network can fully extract the features of small target ships;Secondly,the features of target ship were extracted by convolution neural network(CNN)adaptively,and the features of different levels extracted by the feature extraction network are cascaded to form a multi-level cascade feature extraction network;Then,the target proposal network can find the region of interest in the image according to the features extracted by the convolution neural network,that is,to frame the position of the ship.Finally,through the combination of multiple fully connected layers,the high-dimensional features are mapped into a multi-group,and then the classification function is used to output the probability value of each class of ship,the largest probability value is the class of the ship.At the same time,in order to solve the interference of cloud occlusion and land background,an online hard example mining method is adopted to enhance the learning ability of the network model for "hard" examples with cloud and fog occlusion and land background interference,so to solve the problem of ship recognition under complex sea conditions.The experimental results show that the method adopted in this paper can effectively solve the problem of hard ship recognition under complex sea conditions and the difficulty of recognize small target ships,and realize the recognition of ships under complex sea conditions.At the same time,compared with the basic Faster R-CNN deep learning target recognition algorithm,the precision of the proposed algorithm are improved by 13.6%.The trained model has good generalization ability and robustness,and has good research and application value.