Research On Recursive Feature Instance Segmentation Algorithm For Sea Surface Target Detection

It is of great significance to improve the national marine monitoring system by realizing intelligent detection of sea targets based on remote sensing images and deep learning methods.At present,the detection of sea surface targets is mainly based on ship detection.There are the following problems:(1)There is a lack of ship data in remote sensing.(2)Because remote sensing images are far away from imaging,some ships have small target scale and low pixel ratio,which leads to poor detection results.(3)Because of the variability of ship targets,false alarm and false detection are prone to occur under complicated sea and land background.(4)Due to the influence of imaging mode,the ship distribution in images is sparsity.In view of this,based on the instance segmentation method,the paper studies the detection of ship targets in remote sensing sea surface.The main research contents are as follows:1.In view of the poor detection effect caused by the lack of ship data samples and the small feature content of small targets in remote sensing images,the paper first establishes a fine-grained remote sensing ship target data set with reference to the general dataset MS COCO,which provides strong data support for subsequent research.Then,combined with MS COCO and self-built data set,the small target detection problem is analyzed,and an instance level remote sensing small target data enhancement method is proposed to increase the feature content of small targets in the data set and improve the detection effect of small targets.Finally,small target data enhancement experiments prove the effectiveness of the method.2.Aiming at the problems of small target detection,complex background interference and sparse target distribution in the process of ship target detection in remote sensing sea surface,a two-stage algorithm for ship instance segmentation based on recursive features is proposed.The algorithm is based on Mask R-CNN.Firstly,a recursive feature adaptive fusion network is designed to enhance the feature extraction ability of the model to enhance the sensitivity of the model to small target features.Secondly,the global context enhanced structure is designed to enhance the understanding of the relevance between the target and the environment and enhance the ability of the algorithm to locate the target accurately under complex background.Finally,the Focal Loss is introduced to optimize the classification loss and improve the sample imbalance due to the sparsity of the target distribution.CIOU Loss is introduced to improve the regression loss and further improve the accuracy of ship detection.Experimental results show that the algorithm is effective and feasible in detecting ship targets at sea.3.Aiming at the problem that the performance of the two-stage algorithm is limited by the Anchor mechanism and the lack of inference efficiency,this paper proposes an Anchorfree based single stage ship instance segmentation algorithm based on the recursive feature adaptive fusion network.The algorithm takes PolarMask as the benchmark.Firstly,the recursive feature adaptive fusion network is modified to reduce the amount of computation,and the modified network is used as the backbone network of PolarMask to enhance the feature extraction ability of the model.Secondly,a feature balanced subnet is designed,which balances the low and advanced features of all feature maps by using the balanced semantic features of depth integration and improves the sensitivity of the model to ship scale and shape features.Finally,the edge supervision subnet is designed to improve the problem that the edge blur affects the detection accuracy caused by the irregular shape of the target while not affecting the reasoning efficiency,to further improve the accuracy of the model.Experimental results show that the algorithm is effective and feasible in ship target detection.