Research On Bridge Detection Algorithm Of High Resolution SAR Image Based On Deep Learning

In recent years,with the continuous innovation of SAR technology and the increasing amount of SAR data,high-resolution SAR image has developed rapidly in the field of target detection,and has been widely used in various scenes.Moreover,with the promotion of science and technology,the target detection method based on deep learning has replaced the traditional detection algorithm to become the mainstream.As a vital part of modern transportation facilities,the accurate detection of bridge position has important practical significance in civil and military fields.In this paper,the method of deep learning is used for automatic detection of bridge targets in high-resolution SAR images.The main research work is as follows:1)Aiming at the automatic detection for bridge targets in SAR images,a multi-resolution attention and balance network(MABN)is proposed.Due to the complexity of background information and noise in SAR image,the feature extraction of target is disturbed greatly,and false alarm is easy to occur in the detection results.Therefore,in the feature extraction part,attention balance feature pyramid network is introduced to effectively extract the features of the input image and suppress the interference of background information.Furthermore,IOU balanced sampling is used instead of random sampling to solve the problem of unbalanced sample selection in the training process and improve the discrimination ability of the algorithm.Finally,the classification loss function is improved,and the balanced L1 loss function is used to optimize the training of classification and regression network.In this paper,TerraSAR data with 3-m resolution and Gaofen-3 data with 1-m resolution are used to verify the MABN network,and the detection results are compared with those of Faster R-CNN and SSD.The results show that there are fewer false alarm targets and missed targets in MABN network,the detection rate(P)and average accuracy(AP)are 0.877 and 0.896 respectively,and the recall rate(RR)is 0.917.2)Aiming at the problem of poor real-time performance of bridge target detection algorithm in SAR images based on region proposal network,this paper proposes a one-stage target detection network based on Adaptively Effective Feature Fusion(SSD-AEFF).In order to ensure the detection speed,this paper retains the original backbone network vgg16 and additional feature extraction layer of SSD,but its network structure is simple and the feature extraction ability is insufficient.Therefore,the algorithm proposes an Adaptively Effective Feature Fusion module(AEFF)for three large feature maps to reduce the conflict between adjacent feature layers and extract effective features;The Effective Squeeze-Excitation(eSE)is introduced to enhance the useful features and suppress the redundant features.At the same time,in the loss function,this paper also introduces the Gradient Harmonizing Mechanism(GHM)to solve the problem of sample gradient imbalance and optimize the algorithm parameters in a weighted way.In the experiment,TerraSAR data with 3m resolution is employed,and results are compared with Faster R-CNN,standard SSD and EfficientDet.Results of the experiment show that SSD-AEFF can outperform the other three networks under the condition of ensuring the detection speed,with a detection accuracy of 0.945 and a false alarm rate of 0.103.