Research On Sea Ship Target Detection Method Based On Microwave Active And Passive Fusion

In the microwave remote sensing images of surface ships,the active SAR images contain rich information,but the coherent spot phenomenon is serious,while the passive microwave radiation images contain bright temperature features,but the resolution is low.In this dissertation,we propose a joint convolutional self-encoder based on deep learning technology to generate higher quality fused images for subsequent ship target detection,and also improve the YOLOv3 target detection algorithm to make it more capable of detecting small targets to improve the detection of long-range ship targets.The main research contents of this dissertation are as follows:(1)In this dissertation,an image fusion algorithm based on a multi-branch joint convolutional self-encoder is proposed.The proposed algorithm uses the characteristics of the convolutional self-encoder to reduce the gap between the reconstructed image and the original image as the training goal of the fusion network to ensure the accuracy of the image transformation,while setting different fusion rules by separating the public and private features in the paired images through the joint multi-branching to enhance the rationality of the fusion rules setting.Rationalisation.The fusion network is also designed to address the problems of coherent speckle noise in active SAR images of ships and low resolution in passive microwave radiation images.The final comparison experiments with other algorithms on the dataset used in this dissertation show that the proposed algorithm performs best in three fused image metrics,namely entropy,gradient and mutual information,compared with 5 other image fusion algorithms.(2)This dissertation improves the YOLOv3 algorithm by designing a partially dense residual connection block to replace the residual module in the YOLOv3 backbone network,which increases the gradient diversity and also improves the feature transfer capability of the network.The improved algorithm improves the detection accuracy by an average of12.2% on the multiple datasets used in this dissertation.This dissertation compares 3 singlestage target detection algorithms of the same scale and shows that better detection results are achieved for both the input fused image and the active-passive fused image using the proposed fusion algorithm,and that the best results are achieved when using the proposed target detection algorithm,with a mean average accuracy of The mean average accuracy of0.945 on the simulated active-passive fusion dataset used in this dissertation and 0.913 on the measured active-passive fusion dataset.