Research On Ship Identification And Detection Method Based On YOLOv4

At present,ship detection is widely used in military and civilian fields.In the military field,ship detection can monitor the distribution of ships in key sea areas,detect illegal entry,evaluate the enemy’s combat strength,and form combat intelligence,etc.;in the civilian field,ship detection can be used to monitor surrounding waters in real time,avoid collision accidents,and ensure navigation safety,Assist in disaster relief,manage water transportation,etc.This thesis aims at two different application scenarios of monitoring the distribution of ships in key sea areas and managing water transportation and carries out ship detection research on optical remote sensing images commonly used in the military field and visual images commonly used in the civilian field.Firstly,a visual image-oriented MSD-YOLOv4 ship detection algorithm was proposed to solve the problem of imbalance between ship detection speed and detection accuracy in visual images.Under the premise of ensuring detection accuracy,the detection speed of the algorithm was improved.Based on the YOLOv4 algorithm,the lightweight network MobileNet V3 is used as the backbone network to simplify the backbone feature extraction module.By adding the attention-mechanism module SE to make the network focus on the key information related to the ship,and then get the feature map which is strongly related to the ship features.Inspired by MobileNet series networks,depthwise separable convolution is used to replace partial common convolution and reduce the parameters of network learning.Experimental results show that the proposed visual image-oriented MSD-YOLOv4 ship detection algorithm achieves 45 Frames Per Second(FPS),meeting the requirements of real-time detection,and the m AP(mean Average Precision)of the network reaches 97.8%.Secondly,aiming at the problems of a single scene and limited data types in public optical remote sensing ship datasets,optical remote sensing ship images in various scenarios are collected,and a multi-scene and multi-type optical remote sensing ship dataset named ORShip is established by combining the public datasets HRSC2016 and DOTA-Ship.It lays a foundation for the subsequent optical remote sensing ship detection research.Finally,aiming at the problems of the multi-scale ship,small target aggregation,and complex background in optical remote sensing images,which make it difficult to achieve effective detection and slow model detection speed,a DS-YOLOv4 ship detection algorithm for optical remote sensing images is proposed.Based on the YOLOv4 algorithm,this thesis uses the densely connected network DenseNet as the backbone feature extraction module,which extracts richer ship feature information and improves the utilization rate of features.The attention mechanism module SE is added to strengthen the feature extraction of key information,weaken the influence of irrelevant factors such as complex background,and improve the detection accuracy of the network.Experimental results show that compared with YOLOv4,the detection accuracy of the proposed algorithm reaches 93.44%,which is about 11.3% higher than that of YOLOv4,and the detection speed is also improved.