Research On Infrared Imaging Ship Target Detection Method Based On Deep Learning

Ship is an important carrier of maritime trade,transportation and combat equipment.When war breaks out on the sea,it is easy to become the key target of military firepower.Therefore,ship detection in aerospace remote sensing image has become a research hotspot in the field of remote sensing,and it is widely used in marine safety,fishery management and pollution detection.Therefore,it is particularly important to obtain the ship target location information quickly and accurately in the complex marine environment.The ground-based infrared detector of space-based infrared remote sensing satellite has the advantages of all-weather operation and certain penetrability to clouds and fog.It can realize the ship monitoring in complex marine environment through the difference between the infrared radiation produced by the ship or the reflected solar radiation and the surrounding environment radiation.In recent years,deep learning is a popular method in the field of image processing.The detection model based on deep learning can automatically complete complex feature engineering by neural network,which greatly improves the accuracy and robustness of the network for ship detection in different complex ocean scenes.The infrared remote sensing ship dataset in this paper comes from the infrared ship images obtained from Landsat8 satellite.Aiming at the problems of poor infrared remote sensing image and small ship target.The existing two-stage target detection algorithm Faster RCNN and single-stage target detection algorithm SSD are studied.The main contents of this paper include the following aspects.(1)This paper studies the relevant theory of deep learning convolutional neural network and focuses on the target detection method based on deep learning.The advantages and disadvantages of existing ship detection methods in remote sensing images are analyzed,then to explore the detection method of infrared imaging ship based on deep learning.(2)In view of the lack of infrared remote sensing image ship dataset.In this paper,Landset8 satellite image data is downloaded through the Earth Explorer website,and a ship target dataset of infrared remote sensing image is produced.The data and ship targets in the dataset are marked in VOC data format,and the dataset is divided into training set and testing set.The dataset is expanded by offline image data augmentation.(3)In view of the complex marine environment background,it is difficult to design robust artificial features.The traditional image target detection algorithm training process is cumbersome and the hyper-parameters of the classifier are difficult to set.A ship detection method based on Faster-RCNN multi-scale feature layer is proposed.The improvement of Faster-RCNN includes feature fusion,generating a priori framework more suitable for small targets on multi-scale feature layer and lightweight network design.(4)Aiming at the shortcomings of the two-stage detection algorithm,the detection process is slow and the accuracy is not high enough.By improving the singlestage detection algorithm SSD,this paper proposes a two-step cascade regression ship detection method in infrared remote sensing image combined with visual saliency module.The improvement of SSD includes the introduction of detection proposal idea and visual saliency module,the research of basic feature extraction network and feature fusion method and the realization of real-time accurate detection of small ship target.The method proposed in this paper is verified by experimental data,and the results show that it can realize real-time and accurate detection of ship targets in infrared remote sensing images of complex marine environment.