Research On Instance Segmentation Of Marine Ship Image

In recent years,with the development of deep learning algorithm in the field of computer vision gradually mature,object detection and semantic segmentation technology based on deep learning has gradually penetrated into every field of people’s life,and instance segmentation,as a task of both object detection and semantic segmentation,also has broad application prospects.As an important part of ship intelligent perception and maritime intelligent monitoring system,instance segmentation based on maritime ship image can provide accurate ship position information for the subsequent decision-making of the system and ensure the stable operation of the system,so it has great research value.The specific works are as follows:First of all,since there is no public maritime ship instance segmentation datasets,labelme software was used in this paper to make an instance segmentation label on the collected ship images to construct the maritime ship instance segmentation datasets.The mainstream singlestage and two-stage algorithms were used for performance analysis on the constructed maritime ship instance segmentation dataset,and the single-stage SOLOv2 algorithm and the two-stage Cascade Mask R-CNN algorithm were selected as the baseline algorithm for research.Secondly,the two-stage Cascade Mask R-CNN algorithm was used to carry out maritime ship instance segmentation research.Through visual analysis on the maritime ship instance segmentation dataset,it was found that the bounding box of the Cascade Mask R-CNN algorithm is not accurate and the contour details of the ship are not fully segmented.Therefore,GIoU loss was introduced in this paper to optimize the boundary box and a prediction network of edge and multi-scale information fusion was designed to improve the mask prediction.Through the comparison of visualization and quantitative indicators,it was found that the improved Cascade Mask R-CNN algorithm produces more accurate bounding boxes,more complete instance masks,and better adaptation to ship occlusion scenes.Thirdly,research on instance segmentation of marine ships based on the single-stage SOLOv2 algorithm.Based on the visualization analysis on the maritime instance segmentation dataset,it was found that the SOLOv2 algorithm mis-segmented similar ship instances and could not effectively deal with the interference in the image background.This paper considers that SOLOv2 algorithm is not sufficiently sensitive to the location of ship instances,and then proposes to use spatial attention model to extract the instance distribution information of the classification features and fuse it into the segmentation features to improve the location sensitivity and anti-background interference ability of the SOLOv2.Through the comparison of visualization and quantitative indicators,it was found that the improved SOLOv2 algorithm has excellent instance position sensitivity and anti-interference ability against complex backgrounds.Finally,research on ship instance segmentation for maritime foggy scenes.Foggy ship images of different visibility levels were augmented to construct ship instance segmentation datasets of dense fog and mist fog scenes.And the generalization performance was analyzed on the foggy ship instance segmentation datasets,then we use MobileNetV2 as the scene discriminator,select appropriate model by the scene of the input image to construct a instance segmentation framework which can effectively segment both sunny and foggy scenes.