Deep Learning-based Ship Integrated Detection Method For Remote Sensing Images

With the development of remote sensing image technology and deep learning technology,image processing methods combining remote sensing information and artificial intelligence gradually occupy a major position in the image field,meanwhile,object detection algorithms which using remote sensing images as application object are also emerging.Due to the particularity of remote sensing,accurate and efficient ship detection methods are of great significance for national waters monitoring and territorial security monitoring.However,compared with ordinary images,remote sensing images are easily affected by factors such as light,sea conditions,weather or imaging time,and remote sensing images are characterized by high resolution and rich content.The feature information of remote sensing images is different in varied scenes and environments.For traditional image processing methods,how to correctly process such information is a huge challenge.This paper combines the research contents of object detection and instance segmentation in deep learning,and proposes an integrated detection method for remote sensing image ship detection tasks,which effectively improves the detection accuracy.The main contents of this article are as follows:(1)Introduce the detection method of remote sensing ship image based on traditional image processing,analyze its ability to process remote sensing images in complex scenes and recognition accuracy,and summarize the advantages and disadvantages.Aiming at the defects of using traditional hand-designed features to process images,deep learning is used to improve its accuracy.In this paper,SSD is used as the research object to deal with the task of ship detection,which significantly improves the detection accuracy of ships.At last analyze the reasons for the increase in accuracy and the improved methods used.(2)Collect remote sensing image datasets.The collected data are composed of open source data and manually collected data,and produce label information in a specific format.The overall dataset is divided into a training set,a validation set,and a test set.There are 72860 samples in total and the same amount of labeled information,and the resolution is 768 * 768 px.According to the characteristics of the ship dataset,the SSD framework is improved.The redesigned multi-scale fusion framework,the Focal Loss function that effectively handles the imbalance between positive and negative samples,and the Soft-NMS post-processing method are used to further improve SSD object detection.The accuracy of the ship detection reaches 93.7%.(3)Study the instance segmentation technology based on deep learning,train the collected remote sensing image ship dataset,and add pixel-level labeling information.In this paper,a Mask R-CNN network framework is used to process the ship image,and specific improvements are made to the framework based on the characteristics of the data.The obtained detection accuracy is 92.9%.Analyze the internal connection between object detection and instance segmentation,combine the backbone structures of the two,and make full use of the correlation between the extracted features of the two,so that the prediction frame in object detection and the mask in instance segmentation complement each other to promote and the detection accuracy of the ship was further improved,which finally reached an accuracy rate of 94.3%.(4)Summarize the work of this paper,briefly analyze the speed defects of the neural network-based model,and try to use pruning and quantization to compress the trained model to improve the execution efficiency of the algorithm.It can be found that the compressed neural network has greatly improved compared to the network before optimization,regardless of the amount of parameters and calculations.