Multi-Source Ship Target Fusion Recognition Method Based On Deep Fusion Network

In order to obtain better recognition performance,multi-source fusion recognition combines the information from images provided by multiple sensors.Compared with single-mode recognition,multi-source ship target recognition can achieve higher recognition accuracy and be more robust,thus having broad application prospects in the national economy and military fields.In this paper,a lightweight SAR target recognition network and Dense Net are used as the feature extraction network for SAR images and visible light images.CCA is applied to perform feature transformation and construct a fusion recognition network.The purpose of the study is to improve the recognition accuracy and robustness of the ship target recognition system.The full text of the study is as follows:First of all,this thesis introduces the domestic and foreign research status of SAR target recognition,visible light image recognition and fusion recognition.Then it describes the prospect of the development of related fields.Secondly,this paper studies the application of deep neural networks in the field of SAR target recognition.Considering the difficulty to acquire and label SAR images in large quantities,SAR image data augmentation method is applied before the training of the SAR target recognition network.The method using Gabor filters to fix parameters of shallow layers is proposed to enrich the texture features of SAR images.The structure of fusing different feature vector from different layers is proposed to improve the recognition performance of the network.It achieves recognition rate 99.71% on the MSTAR data set and recognition rate 99.21% on the ship SAR image data set.This paper also applies the Dense Net-49 structure as the visible light image recognition network,which greatly reduces the number of network layers and the total number of parameters without loss of recognition accuracy on the base of Dense Net-121 in ship visible light image recognition task.The network achieved recognition rate 99.35% on the ship visible light image data set.Finally,this paper studies the feature fusion method based on CCA.Our method transforms the features of heterogeneous image features through CCA,and designs a classifier network composed of fully connected layers.At the same time,the amount of visible light image and SAR image data which are one-to-one matching is small.In contrast,the amount of the visible light image data set and the SAR image data set is relatively much larger.This paper proposes to train single-mode recognition neural networks by using visible light data set and SAR data set,and then use the matched data to train the classifier of fusion network.The experimental results have a significant increasement on recognition accuracy with noise-added images,while comparing with the recognition accuracy of the single-mode recognition network.