Ship Detection Algorithm Of Remote Sensing Image Based On Deep Learning

With the increasing demand for marine resources,there are more and more ships appearing on the sea.It is particularly important to monitor and manage those ships on the sea.The thesis does researches on the ship detection of the large-scale sea surface remote sensing image,and proposes a two-stage ship object detection algorithm which includes pre-classifiers and detectors.The algorithm has the characteristics of high accuracy and fast speed,which provides strong support for the realization of marine monitoring.The main work of the thesis is as follows:Firstly,unlike general object detection tasks,large-scale remote sensing image object detection has the characteristics of large image size,small target and large-scale range,serious background interference and large amount of calculation.A two-stage large-scale remote sensing image object detection algorithm,which includes pre-classifiers and detectors was proposed in the thesis.Using the “divide and conquer” idea,the large-scale remote sensing image is divided into sub-graph processing.The algorithm includes four steps: sub-graph sliding window slice,sub-graph pre-classification,sub-graph detection and results merging.Experimental results show that the method can complete the task much more quickly and accurately.Secondly,considering that the sea background is often covered under the cloud,the thesis construct a large-scale remote sensing image ship detection dataset.In the processing of large-scale remote sensing images,in order to ensure the integrity of ship,we design a slice method to ensure a certain overlap rate.At the aspect of sub-graph pre-classification algorithm,the thesis studies and designs a spatial position weighted module based on selfattention mechanism.The experiments show that this module can effectively improve the recall rate.Thirdly,for the sample imbalance problem in object detection algorithms,the thesis proposes an object detection algorithm based on sample loss equalization.The algorithm conducts research from two aspects: classification and regression loss equilibrium and network parameter setting.In addition to this we also designs the object detection network based on anchor.The experiments show that this method can get much more accurate detection results than other similar methods.Finally,although to solve the object scale problems we can usually introduce the feature pyramid network into the object detection algorithm,but it will also lead to the problem of mismatch between the object scale and receptive field.This thesis constructs an object detection algorithm based on the equilibrium of layer loss in feature pyramid network.The algorithm balances the loss among layers to ensure that the parameters of each layer can be fully trained.Experimental results show that the method achieves much higher accuracy detection results.