Research On Ship Target Detection And Classification Method Based On Deep Learning In High Resolution Optical Remote Sensing Images

China has a vast sea area.In civil and military fields,high-resolution visible light image ship target detection research has a very high degree of attention.Its technology can be widely used to monitor shipping traffic,protect marine rights and interests,and improve coastal defense early warning capabilities.Wait for the scene.Therefore,in recent years,many excellent research results have emerged in the research on ship target recognition for high-resolution visible light images.These research methods are mainly divided into traditional sliding window search detection methods and manual design feature combination methods and depth-based methods.Integrated detection and recognition method of convolutional neural network model.The former learns more about shallow semantic image information features.Through human eye observation,it can design specific texture feature extraction modules for different application scenarios to meet the needs of image detection and classification,but this method also There are certain drawbacks.The hand-designed feature extraction module is often simpler than the latter and does not provide good image detail information;the latter can learn rich image texture information and combined features of different semantic levels through deep convolutional neural networks.,Not only is more effective for the detection task,but also plays a very good role in identifying the fine classification task of the ship ’s target.Nevertheless,there are still many challenges in the direct application of deep convolutional neural network applications and remote sensing image scenes.The analysis in this paper mainly has the following difficulties: 1)In large-scale satellite remote sensing images,the ocean background,especially the port background,is more complicated.Ship targets are easily confused with the land background.At the same time,the scale of different ships is very different,and it is difficult for the model to take into account targets of various sizes;2)In actual use,satellite images are viewed from above,and the ship targets are long and narrow and have many directions In the detection process,the vertical frame labeling will cause more background noise in the detection frame,and the oblique frame labeling needs to be used to solve this problem;3)There are many types of ship target subdivisions,and the differences between the categories are small.Mining It is very difficult to express the characteristic information of different types of ships to achieve a good classification effect.In order to solve the above problems,this subject has carried out research on ship target detection and classification methods based on deep learning high-resolution satellite remote sensing images.In this paper,based on the improvement of ship target detection accuracy,ship direction detection and ship fine classification accuracy,related research on target detection and target classification algorithm based on deep learning is carried out.main tasks as follows:1.Analyze and summarize the classic convolutional neural network model and training detection process based on convolutional neural network remote sensing image target detection,and also make analysis and summary of the application of classic convolutional neural network classification model in ship targets Evaluation;2.Aiming at the problem of large differences in ship scale distribution and variable directions in port images taken by remote sensing satellites with complex backgrounds,a cascaded convolutional neural network based on multi-layer feature fusion and a Canny-Hough ship direction detector Ship target detection method.Among them,the Canny-Hough ship direction detector module aims to return the direction detection frame from the vertical detection frame with more complex background.The framework first uses a multi-layer feature fusion area convolution extraction network to extract target candidate regions;then uses a three-layer cascaded target detector to progressively train and detect ship targets,and the final vertical detection is obtained after three-layer detection Candidate box;in the next processing,the ship detection slice is passed through the Canny-Hough ship direction regressor to obtain the final ship detection area.After the effectiveness test and experimental verification on the data set collected in this experiment,the method in this paper can effectively detect ship targets in different directions.3.In view of the difficulty of fine-grained classification of ship targets,this paper proposes a ship target classification model based on Res Net network and triple loss function.This paper studies the existing fine-grained classification methods of deep learning.The feature representation is compressed while improving the use of the triple loss function as the loss measurement standard.In this paper,the framework first preprocesses the image to correct the placement of the ship target,and then uses the Res Net network as the image feature extractor,and The feature vector of the image is mapped to a 256-dimensional vector space,and then the triplet is calculated to train the calculation loss.Compared with the classical convolutional neural network classification model and the fine-grained classification method,this method has certain advantages when applied to ship targets.The classification experiments conducted on the HRSC 2016 dataset verify the effectiveness of this method.