Research On SAR Target Detection Method Based On Improved YOLOv4

In recent years,with the rising trend of ocean globalization,the strategic importance of ship targets has been increasing.As the main carrier of all kinds of activities on the sea,rapid and accurate identification of ship targets has always been the focal point of research in the field of identification and identification of targets.With the development of synthetic aperture radar(SAR)imaging technology,its anti-jamming ability and sustainable observation are becoming more and more obvious.At the same time,more and more SAR ship images can be used for ship target detection and recognition.Currently,the continuous development of Convolutional Neural Network(CNN)has achieved good results in both target detection and target recognition.Under the background that the Convolutional Neural Network has a large number of data samples,how to use a large number of SAR images efficiently to improve the accuracy of ship target detection is a difficult problem in ship target detection.After comparing the current mainstream target detection algorithms,it is found that YOLOv4 has a balanced target detection accuracy and speed,and does not require high hardware conditions.For this reason,this paper focuses on how to improve the accuracy of YOLOv4 algorithm for SAR ship target detection.Based on YOLOv4 algorithm,this paper studies SAR ship detection.The main work includes:(1)The history,characteristics and concepts of the principle of photographing synthetic aperture radar are introduced.The status of SAR vessel image target detection investigation shall be analysed and summarised.This work mainly introduces the network structure and development process of convolution network and introduces in detail a representative singlephase target detection algorithm and a two-phase target detection algorithm.(2)A YOLOv4 enhancement algorithm is presented which fuses the receptive field module.In the target detection algorithm,the inadequate acquisition ability of the sensing field will directly affect the accuracy of the algorithm to detect the target.For this reason,this paper proposes using the sensing field module to improve the acquisition ability of YOLOv4 algorithm,using multi-branch structure to obtain different size of target features,and then effectively enhancing the sensing field on each branch through void convolution to get more rich feature information.Finally,experiments show that the average accuracy(m AP)of this method on SAR-Ship-Dataset dataset is 94.54%,1.19% higher than the original YOLOv4,and the m AP value is better than other mainstream target detection algorithms.(3)A SAR ship detection algorithm based on anchor frame scale transformation and attention mechanism is presented.The characteristics of SAR imaging cause unavoidable noise in its image,which affects the detection and recognition of targets by the algorithm.At the same time,whether the initial anchor frame setting of the algorithm is suitable for training datasets also affects the detection accuracy of the algorithm.To solve the above problems,this paper uses the attention module to reduce the impact of background noise on the detection accuracy of the algorithm,and uses the improved K-means clustering algorithm to fit an anchor frame suitable for the dataset.Then,a set of ablation experiments is used to verify the validity of the two parts.Finally,experiments show that the average accuracy(m AP)of the method is better than other algorithms on SAR-Ship-Dataset dataset.