Research On SAR Target Detection Method Based On Heterogeneous Auxiliary Information

Remote sensing technology uses various types of sensors on multiple platforms to obtain information in areas of interest through non-contact means.They have been widely used and provided important information support,such as resource and environmental surveying,crop growth monitoring,natural disaster early warning,urban development monitoring,post-disaster assessment and reconstruction guidance,etc.Synthetic Aperture Radar(SAR)has all-time and all-weather characteristics.So SAR become an important composition to maximize the information obtained in multi-sources.Nowadays,new challenges followed by the development of imaging technologies as well as the improvement of spatial resolution and target details.On one hand,the increasing size of image and the complexity of the features have weakened the salience of the target area and even the target itself.How can the traditional method of using coarse threshold detection and filtering to remove false alarms,achieving complete extraction of the target in a larger area and clearer background? On the other hand,the development of machine learning technology provides an end-to-end solution for SAR image target detection tasks.But the problem of limited samples in the network training process is unavoidable.How to take the target as the traction and make full use of the the auxiliary information of heterogeneous and heterogeneous data is an urgent problem should be solved in network learning under the condition of insufficient samples in the SAR domain.This dissertation firstly performed pre-processing before target detection from two aspects: target region extraction and training sample selection.Then independent methodological researches are conducted on the three application situations based on target geographic location or the availability of prior information.An association model to unify heterogeneous feature expression was established in this thesis,and it completed the integrated process of information reuse through transfer learning network.The means of association and migration in this thesis ealize the "positive" supplement of heterogeneous complementary information and the distinguish of target and background.This dissertation breaks through the mindset of traditional multi-source collaborative processing,and a new idea for SAR target detection based on target-driven fusion of multi-source auxiliary information is proposed.Firstly,for the first case,when the prior information is extremely limited,traditional detection methods provide the only possibility for target detection tasks.In this case,the focus is how to configure different detection and post-processing strategies according to target characteristics.So,it is possible to balance target extraction and interference rejection.For example,for large-scale targets represented by airplanes and ships,a multi-layer constant false alarm detection method with adaptive estimation of the background model was proposed.For near-shore ships which are easily affected by ground objectives,the false alarms in the detection process are extremely high.A detection method taking into account both the local and global saliency factors to obtain the saliency information of the target was proposed.Then,aiming at the problem of insufficient target domain samples in the training step when using machine learning methods,this thesis discusses how to use target association and information transfer technology to achieve optical information assistance for different application backgrounds.For the second case,when the number of prior samples of the auxiliary domain can meet the requirements,the biggest problem is the difference between heterogeneous samples.Aiming at the problem that heterogeneous samples from multiple sources having large differences in feature space and cannot be unified,a target-oriented multi-source association model was established.This model broke through the dependence of the pixel-oriented multi-source fusion algorithm on the registration results.It was drived by targets to realize the unified representation of heterogeneous samples at the feature-level.Firstly,a bridge between heterogeneous samples was constructeda triple-uints-graph model including samples,features and labels.This model weakened the differences between heterogeneous sources,and the consistency of samples in the same category were used at the pixel and feature level layer-by-layer.Then,the feature constraints were introduced into the representation learning process of the graph model.A unified feature space for the feature expression of heterogeneous samples was generated.Experimental results showed that the use of associated features from the source domain(optical)can effectively enrich the target domain(SAR)prior feature set.It was effective to assist the completion of the target detection task when the quantities of target sample is insufficient or unbalanced.Finally,for the third case,the lack of prior samples makes most methods invalid.Aiming at the problem of insufficient prior information and the inability to obtain a stable high-performance network,this thesis was oriented to the task of target detection.Under the framework of convolutional neural network(CNN),a transfer learning model trained by optical-assisted sample constraints was constructed.This model further reduced the dependence of network training on prior samples,and realized online testing of target domain samples.At first,a stable CNN model trainded on the standard natural data set can extract general graphic features from a variety of inputs.Then,a fine-tuning and transfer parameter transfer CNN was constraint trained on the optically-assisted slices,so as to specifically adapt to different tasks.Meanwhile,a feature association method based on morphological processing improves the feature difference brought by different sensors for training and testing samples.It was achieved through the transformation of heterogeneous image in pixel level.Experimental results showed that the target detection rate can reach 90% when there is no prior sample in the target domain and a few prior samples in auxiliary domain.