Ship detection method based on attentional guidance and multi-sample decision
The ones-stage object detection method has the characteristics of fast training speed and short inference time.However,its feature pyramid network(FPN)cannot suppress the background and noise information of the synthetic aperture radar(SAR)ship image,and the detection head has a prediction bias.This paper proposes a detection model based on attention guidance and multi-sample decisions for SAR ship detection.Firstly,in order to improve feature representation,this study suggests adding an attentional guidance network to the top of the feature pyramid in order to decrease noise and background interference.Secondly,Multi-sample decision networks are proposed to participate in predicting ship locations.By increasing the amount of output samples in regression branches,the network reduces the impact of prediction bias on detection outcomes.Finally,a novel maximum likelihood loss function is designed.The loss function constructs the maximum likelihood function from the output samples of multiple decision networks,which is used to standardize the training of decision networks and further improve the accuracy of target positioning.Compared with RetinaNet and current advanced object detection methods,the proposed method shows higher detection accuracy on the SSDD dataset,with AP up to 54%.Compared with the baseline method,the SARetinaNet method improved the AP evaluation index by 3.4%～5.7%,the number of training parameters Params only increased by 2.03M,and the FPS only increased by 0.5iter/s.
ship detectionattentional guidancemulti-sample decisionmaximum likelihood loss functionone-stage detectionsynthetic aperture radar
万方数据
维普
