Research On Boat Recognition Method Based On Deep Convolution Neural Network In Complex Sea Surface Environment

The high level of ship intelligence not only helps to improve the automatic law enforcement ability of the maritime surveillance department,but also plays a great role in the regulation of marine resources and the defense of land security.As one of the core technologies of the intelligent ship,ship identification technology provides more effective decision-making information through the realization of objective understanding and attribute analysis,which lays the foundation for the autonomous obstacle avoidance and autonomous decision-making of the intelligent ship,thus improving the intelligent level and safety guarantee ability of the ship.Based on the above problems,a ship identification method based on deep convolution neural network in complex sea surface environment is proposed.The main contents of the study are as follows:(1)For some ship samples with low target and background differentiation,noise interference and rain fog occlusion,the model cannot extract more effective feature information,resulting in low accuracy of ship identification.In this paper,the image enhancement algorithm,the improved weighted joint filtering algorithm and the improved dark channel prior defogging algorithm are proposed to preprocess the test samples to improve the discrimination between the object and background of the samples and to reduce the influence of external environmental factors on the ship identification task,thus improving the accuracy of ship identification..(2)Ship identification is one of the most challenging tasks in maritime traffic monitoring,which requires accurate classification and positioning identification of very small-scale ship targets in complex sea surface scenarios.In this paper,by studying the theory and algorithm of deep learning,combined with the requirements of high accuracy and real-time of ship identification task,an improved fast regression detection algorithm,YOLO-G is proposed.by reconstructing the underlying network Darknet-65,the algorithm extracts the features of the ship target and uses the multi-scale feature interaction layer to highly fuse the shallow position information and the deep semantic information to obtain the coordinate information and the category label of the target,and selects the prior box mechanism and the modulation loss function to optimize the network model to further reduce the interference of the invalid feature information to the recognition accuracy and the model accuracy of the minimal target.(3)In this paper,the self-built ship dataset and MS-COCO dataset are selected to test and evaluate the YOLO-G network model to verify its effectiveness.The experimental results show that the improved YOLO-G model has achieved 96.5% recognition accuracy in the classification task of ships,which is 6.5% higher than the YOLOv3 algorithm,and realizes the correct classification of ship targets in the complex recognition environment.in terms of model accuracy,YOLO-G obtains 34.9% average precision mean(map)on the MS-COCO dataset.its model accuracy is much higher than that of advanced single-stage detectors such as SSD,DSSD,YOLOv2,YOLOv3 and so on.compared with YOLOv3 model accuracy,the accuracy is improved by 5.7%,which meets the task requirements of complex sea surface ship identification.it is proved that the proposed ship identification algorithm is feasible and has high research and application value..