Ship Target Detection In The Optical Remote Sensing Image Based On Deep Learning

With the improvement of the resolution of optical remote sensing images,the detection and recognition of key targets on remote sensing images has important research value and practical significance.As an important transport carrier of the sea and a key military target for wartime strikes,ships have their broad application prospects in the civil and military fields.The traditional sea ship target detection algorithm is difficult to meet the needs of ship intelligent detection due to the influence of sea surface uncertainty such as illumination,fog,ship distribution and target distance on complex ocean background.The deep sea-ship target detection algorithm based on deep learning mainly relies on the unique advantages of neural network in the learning of big data features.It can quickly and effectively extract the representative and distinctive features of the target from the mass data in the form of hierarchical learning.However,there are still many shortcomings in the existing research on sea-ship detection based on deep learning.1)There is no systematic classification test for ships.2)The data sample size of the trained neural network model is too small,resulting in model robustness.Poor 3)The neural network model used for ship detection has not yet reached the balance between detection accuracy and detection speed.The detection effect is mostly high speed,low precision,high precision and low speed.In view of the above deficiencies,this study established a sample library of the system’s ship classification,and used and improved the YOLOv3 network to realize real-time classification and detection of marine ship targets on optical remote sensing images.The main work of this paper is summarized as follows:(1)A ship classification data set for high-resolution remote sensing images is established.A marine ship classification test data set containing more than 14,000 remote sensing images and a total of approximately 24,000 ships of different types was established.Firstly,based on the classification rules of ships in the "Regulations on Ship Registration" of the Maritime Safety Administration of the People’s Republic of China,combined with the actual situation of remote sensing images,a basic classification system for marine vessels based on remote sensing images was established.Elected nine types of ships such as "passenger ship","bulk ship","container ship","liquid cargo ship","engineering ship","barge","tugboat","yacht" and "military vessel" Part of the classification system.The high-resolution optical remote sensing image is selected for data preprocessing,and then the ship target on the remote sensing image is marked according to the classification system,and the data is cleaned during the marking process to eliminate the noise data.Secondly,according to the training characteristics of neural network,the data set is formatted and unified to meet the requirements of deep learning for input data format.(2)Established a deep learning training platform,using the YOLOv3 and Faster R-CNN neural network algorithms with outstanding detection results in the target detection field to classify and detect the established marine ship dataset.The trained model is validated on the test set.The detection and detection rate of YOLOv3 is significantly better than Faster R-CNN.Although the YOLOv3 network has a good overall detection effect on the ship classification test set,the test results for the small target type ship are still not ideal.An improved network based on YOLOv3 was designed for this problem,and the training process of the network was improved at the same time.The improved YOLOv3 network increases the three scale detections of the original YOLOv3 network to four,and uses K-means clustering algorithm to cluster the image target size in the training samples to generate a unique anchor of the ship classification data set.Point frame parameters,the clustering results are directly applied to the training and detection of the network to improve the final detection capability of the network.In the network pre-training process,remote sensing ship images with different resolutions are used as the network input fine-tuning network.In the network iterative process,the size of the input image is randomly changed,and multi-scale network training is performed to make the model robust to the detection of images of different sizes.Tests show that based on the improved network of YOLOv3,the accuracy and recall rate on the ship classification test set are better than the original YOLOv3 network,especially in the detection of small target type ships.(3)Build a ship detection system in the Linux platform,set up a comparison experiment,and compare the improved YOLOv3 network with YOLOv3,Faster R-CNN,YOLOv2,YOLOv2-tiny,and YOLOv3-tiny in the ship classification test set.Detect the effect.Among them,the improvement of YOLOv3 detection is the best,the recall rate reaches 95.62%,and the flatness accuracy rate reaches 93.89%.According to the test results,the reasons for the different detection effects of each network are systematically analyzed,and the respective advantages of each network in detection rate and classification and positioning are clarified.