Saliency-aware Convolutional Neural Network For Ship Detection In Surveillance Video

In order to be able to learn about the state of marine resources and ensure the safety of the ocean,mankind urgently needs to accurately monitor the state of the ocean.How to detect the types and positions of ships in a timely,efficient,low-cost and automated manner has great application value and research significance for marine security work.With the continuous development of computer vision technology,ship target detection based on visible light images has become a hotspot in the field of surveillance video,which has attracted widespread attention from scholars at home and abroad.The great achievements of CNN in natural image classification and target detection have led many scholars at home and abroad to abandon the traditional methods based on artificial predefined features and choose convolutional neural networks to conduct research on ship target detection.And salient target detection can help people quickly locate the target area of interest in the image.It is widely used in computer vision related tasks.Inspired by these efforts,many researchers have begun to use significant information in vessel inspections.The paper first introduces the basic theoretical models and methods of ship target detection and saliency detection.First introduce the detection process of the traditional ship detection algorithm,then introduce the common target detection models and algorithms based on CNN,then introduce the common saliency detection models,and discuss the saliency detection algorithm from the perspective of local contrast and global contrast.Finally,the experimental evaluation index of ship target detection algorithm is introduced.Secondly,the paper combines a deep learning algorithm and a saliency detection algorithm,and proposes a convolutional neural network(Saliency-aware CNN)that combines saliency features and coastlines.For the situation that convolutional neural network is easy to misclassify,it extracts reliable coastline features to assist the convolutional neural network classification;for the case where the convolutional neural network ship position prediction is not accurate enough,the area frame generated by the convolutional neural network is combined with coastline features Carry out saliency detection,optimize the final ship position,and effectively improve the results of ship detection.Third,this paper conducts experiments based on the Seaships ship dataset created in the group.The experiments are compared from five aspects:(1)the Saliency-aware CNN method proposed in this paper is compared with the mainstream target detection methods to verify the accuracy and real-time improvement of the algorithm,and the experimental results are summarized;Other saliency detection methods combined with convolutional neural network are compared with this algorithm to verify the advantages of the saliency detection method selected in this paper and summarize the experimental results;(3)the algorithm proposed in this paper and the algorithm without added coastline features Make comparisons to verify the role of coastline features;(4)then compare the algorithm proposed in this paper with traditional ship detection algorithms to prove that the algorithm has higher accuracy and real-time performance in complex environments;(5)in order to By comparing the effects of surveillance video vessel detection and remote sensing vessel detection,the ship detection method of satellite remote sensing image is realized,which proves the advantages of the algorithm proposed in this paper in real-time vessel detection.The main innovations of this paper are:Based on YOLOv2 algorithm,a convolutional neural network(Saliency-aware CNN)combining saliency features and coastline features is proposed.First use the convolutional neural network to predict the class and initial position of the ship,and assist the convolutional neural network classification by extracting reliable coastline features;then extract the salient features of the area frame generated by the convolutional neural network,and combine the coastline features in the salient During the performance testing,the final ship position is optimized to make the final ship position more accurate.