| Intelligent detection and recognition of marine targets is the basis of naval battlefield situation assessment and threat estimation.The core task of intelligent detection and recognition of ship targets in optical remote sensing images is to determine whether there are ship targets in the images or not,and we need to detect,recognize and locate the ship targets accurately.It has a broad application prospect in fishery management,maritime rescue,maritime traffic monitoring and sea battlefield situation awareness.However,due to the large amount of aerial image data and the fact that aerial images are susceptible to various factors such as sea surface conditions,weather conditions,illumination conditions,and imaging detector parameters.There are fogs,sea clutter,clouds and other natural environment interference on the sea surface,which makes the image blurred and the signal-to-noise ratio of the image is very low.The time-varying sea environment interference also makes it difficult to suppress the background noise.These environmental interferences are likely to cause false alarms and missed detections of ship targets.The shooting distance of aerial image is long,and the target occupies only a small part of the entire image,so it is difficult to identify the ship targets with small scale and few features.Different shooting angles,different shooting distances and different ship orientations make ship targets in aerial images have variability in scale and rotation,which have a great impact on the accuracy of detection and recognition results.The thesis focuses on the situation that the airborne photoelectric reconnaissance system relies on the operator to manually interpret ship targets based on the displayed images during the work process,which leads to slow interpretation speed and easy to be affected by subjective factors.This paper aims to improve the intelligence level of sea reconnaissance information processing,focusing on the problem of intelligent detection and recognition of ship targets in optical remote sensing images,carry out research on key technologies such as ship target feature extraction,feature fusion,target detection and recognition.Our goal is to achieve high-precision detection and recognition of ship targets on the sea and improve the probability and accuracy of ship recognition,and provide a basis for situational awareness of maritime targets.The main research contents of this article are as follows:1.In order to improve the accuracy of ship target detection and recognition in remote sensing images,data preprocessing methods are studied.The ship image data obtained by aerial reconnaissance is limited.The affine transformation is used to change the perspective of ship targets in satellite remote sensing images to make up for the problem of insufficient aviation ship target samples in military scene,and to improve the accuracy of ship detection and recognition.Atmospheric absorption and scattering affect the contrast and clarity of remote sensing images acquired by optical system.The remote sensing images under sea scenes are corrected by using the theory based on dark channel prior,which reduces the influence of the absorption and scattering of light by water vapor and particles in the atmosphere on the quality of remote sensing images.It further improves the accuracy of ship target detection and recognition in remote sensing images.2.In order to improve the accuracy of single-stage target detector for ship target detection and recognition,an improved YOLOv3 model of ship target detection and recognition algorithm is proposed.For the problem of small target size and large scale change in ship target detection and recognition,the design of the IOU threshold has been improved to increase the recall rate of ship target detection.For the problem of large attitude change of ship target in the image,the model is further optimized by the method of multi-image hybrid data enhancement and the balanced distribution of the ship target angle.These methods effectively improve the accuracy of ship target detection and recognition.Experimental verification shows that the algorithm has achieved good results in the detection and recognition of ship targets.3.We study the network model of ship target detection and recognition in remote sensing images based on convolution feature fusion.In order to improve the accuracy of multi-scale ship target detection and recognition,the target detection and recognition network structure of convolutional feature fusion was designed.Through the fusion of different levels of feature maps in the convolutional neural network,the semantic information from the deep layer and the fine-grained information from the shallow layer are combined.The multi-scale fusion feature map is used for ship target detection and recognition,which improves the adaptability to large,medium and small ship detection and recognition.Furthermore,combined with the anchor frame design,the accuracy of ship targets detection and recognition is improved,especially the detection and recognition effect of small-scale ship targets in remote sensing image is enhanced.In order to improve the real-time performance of the model,pruning strategy is adopted to compress the model,which reduces the storage space required by the model and improves the detection speed of the model.To sum up,the paper analyzes the difficulties existing in the intelligent detection and recognition of ship targets in optical remote sensing images.And we have conducted research on related theories.Based on the convolution neural network model of deep learning,a deep learning framework suitable for the detection and recognition of ship targets has been built,and we have achieved certain research results.The research results of the thesis provide a theoretical basis for the intelligent detection and recognition technology of ship targets in optical remote sensing images,and the research of this paper has reference significance. |
