Marine Ships Object Detection Based On Deep Learning

Intelligence is the main direction of ship development and one of the key development areas of "Made in China 2025".Smart ships have a huge application space in both civil and military ships.One of the core goals of smart ships is to navigate safely and autonomously,which puts forward a higher level of perception of the surrounding environment during the course of the ship’s navigation.With the gradual application of high-resolution visual perception systems on ships,vision-based maritime object detection and recognition technology has received more and more attention.With the rapid development of deep learning,the ship object detection technology based on deep learning occupies an important position in the key technology of intelligent ship autonomous navigation situation awareness based on its classification and positioning tasks of the object.This paper focuses on the marine ship object detection algorithm based on deep learning,and aims to solve the problems of ship object detection in marine scenes.The main work is as follows:First,a dataset of marine ship object detection was constructed.The dataset has 6303 ship images in marine scenes,with a total of 21,333 labeled objects,including six types of objects:liner,container ship,bulk carrier,island reef,sailboat,and other ship,from background,object scale distribution,meteorological conditions,light conditions,etc.The data diversity design is carried out from the perspective.Compared with other existing ship object detection datasets,this dataset has significant advantages in data richness and scientific evaluation.Second,the performance analysis of mainstream object detection algorithms in maritime scenes is carried out.According to the marine object detection dataset constructed in this paper,the performance of the current mainstream object detection algorithms based on deep learning applied to marine object detection is compared and analyzed from three aspects:average accuracy,category average accuracy,and speed through experiments.The results show that all algorithms are There is a problem that the detection effect of small-scale ship objects and ship objects in foggy scenes is not good.Therefore,the Cascade RCNN algorithm with the highest detection accuracy is selected to improve,and these two problems will be solved in the followup work.Third,the research on the object detection algorithm for small-scale ships at sea has been carried out.Aiming at the problem of low detection accuracy of marine small-scale ship objects due to limited feature information,an improved Cascade RCNN-based marine small-scale ship object detection algorithm is proposed.The improved algorithm introduces switchable cavity convolution into the backbone network,and at the same time adds feature fusion channels to the feature pyramid network and uses sub-pixel convolution for upsampling,thereby achieving the improvement of the detection accuracy of small-scale ship objects in marine scenes.Fourth,the research on the ship object detection algorithm in the foggy sea scene was carried out.A marine foggy ship object detection dataset is constructed,and the influence of sea fog on the detection results of maritime ships is analyzed through experiments.Based on the analysis results,a foggy maritime ship object detection algorithm based on improved Cascade RCNN is proposed.The improved algorithm uses boundary-aware positioning in the cascade detector and deformable convolution in the backbone network to improve the accuracy of ship object detection in the foggy sea scene.