Research On Deep Learning-based SAR Object Classification And Detection Technology

Benefiting from its ability that can work in all weather,all day and all night,synthetic aperture radar(SAR)has been widely applied in reconnaissance,detection,geological prospecting,disaster detection,target detection,public security check,etc.As the basic problems in SAR image analysis and image parsing,SAR object classification and detection attract more and more attention and become significant.Because deep learning has made great success in computer vision,in this dissertation,the deep learning-based SAR object classification and detection methods are researched.Recently,deep learning-based methods have dramatically improved the performance of object classification and detection.In this dissertation,SAR ground target classification,SAR ship classification,SAR ship detection,and SAR security foreign detection are considered.Aiming at the problem that deep supervised networks rely on large-scale labeled data and considering the property of the distribution of SAR objects,the research of deep learning-based SAR object classification and detection on network design and training with inexplicit-labeled data is carried out in this dissertation.The main contributions are as follows.(1)The influence of inexplicit-labeled data on network training is analyzed,and that the deep network relies on a large-scale dataset is pointed out.The inexplicit-labeled data include limited labeled data and noisy-labeled data.In this dissertation,research on training the classification network with limited-labeled data and noisy-labeled data is carried out.Aiming at the limited performance with limited labeled data,a batch balance and consistent augmentation-based semi-supervised framework is proposed in this dissertation,which allows a SAR classification network to utilize unlabeled data during training and ultimately alleviates the demand of labeled data.The proposed batch balance can ensure the prominent effect of the supervised learning part of the framework for training by balancing amounts of labeled and unlabeled samples in a mini-batch.Unified augmentation can ensure the labeled samples and unlabeled samples share the same augmentation method,which allows the network to learn the transform-invariant feature.The experiments on multiple SAR object datasets prove the effectiveness of the proposed semisupervised framework on improving the classification performance with limited labeled data.The influence of noisy-labeled data on classification network training is also analyzed.Aiming at the influence of noisy labels,a loss curve fitting-based method is proposed in this dissertation,which can identify the noisy labels and train the SAR classification network effectively.Label noise modeling is achieved by unsupervised clustering via fitting loss curves to identify whether the sample's label is clean or noisy.Then the network is trained using augmented samples with clean labels to correct noisy labels further.The experiments on the SAR dataset demonstrate that the proposed method can correct 97.9% of noisy labels effectively and obtain 99.2% classification accuracy with40% noisy-labeled data.(2)The design of object detection networks based on the property of the distribution of SAR objects is researched,in which SAR ship detection and SAR security foreign object detection are included.In this dissertation,different properties of the distribution of SAR objects are considered.Based on the regularity of the distribution of SAR security foreign objects,a normalized accumulation map-based millimeter-wave foreign object detection network is proposed.The proposed normalized accumulation map(NAM),calculated as the average of binary masks representing the object location for each image,can reveal the positions of frequently-appeared concealed objects.NAM offers different weights for different locations when computing confidence loss.Experiments on a millimeter-wave foreign object dataset demonstrate that the proposed normalized accumulation map-based training mechanism effectively improves the mean average precision by 4.43%.The proposed NAM is easy to implement for object detection networks since only the confidence branch is related to the normalized accumulation map.At the same time,it keeps the loss scale stable compared to the original loss and does not need to adjust the weights between regression loss,confidence loss,and classification loss.Aiming at the problems of complex anchor settings caused by the variety of the distribution of SAR ship,an anchor-free and sliding-window-free deconvolutional region proposal network is proposed,which avoids the difficulty of parameter tuning caused by anchor mechanism and can generate accurate and concise reference boxes.Furthermore,aiming at the problem that the deconvolutional neural network is not sensitive to closeobject contour,a center-ness-based anchor-free SAR object detection is researched to improve the detection performance further.The experiments on a SAR ship dataset prove that the researched anchor-free methods can simultaneously generate reference boxes effectively and achieve accurate object detection.(3)The training methods for SAR object detection networks are researched considering complex detection annotations.The complexity of detection annotations causes more difficulty in training the object detection network with inexplicit-labeled data,in terms of difficulty in label assignment for unlabeled samples and complicated label noise.Based on the detection network design,the semisupervised framework for classification network,and the classification label noise modeling and correction in the previous chapters,SAR object detection network training under complex detection annotation with limited-labeled data and noisy labeled data is researched,respectively.For training with limited-labeled data,in this dissertation,considering the complexity of detection annotations,a semisupervised learning-based SAR object detection training method combining label propagation and consistency constraint is researched,which can utilize unlabeled samples during training and thus improve the performance of SAR object detection network effectively.For training with noisy-labeled data,in this dissertation,a region label noise modeling and correction-based SAR object detection network training method with multiple kinds of noisy labels is researched,which corrects the noisy labels during training,weakens the influence of noisy labels,and thus improves the detection accuracy of SAR object detection network with noisy labels.In this dissertation,when applying deep learning to SAR object classification and detection,the problems that deep networks rely on large-scale labeled data and different SAR objects have different properties of the SAR object distribution are considered.Aiming at the problem,the semisupervised learning framework,the noisy label correction method,and the network design are researched,respectively,to improve the performance of classification and detection networks.The above research on deep learning-based SAR object classification and detection technology is significant for SAR image analysis and image parsing.