master/repos/SAR_specific_models-master

• SAR_specific_models
• • Environment
  • SAR-Specific Models
  • • Target Detection
  • References
  • Deep SAR-Net (DSN)
  • • Training Procedure
    • Testing Procedure

SAR_specific_models

1. This project provides some SAR specific models with strong abilities to extract spatial features of single-polarization Synthetic Aperture Radar (SAR) amplitude images.

2. A novel deep learning framework Deep SAR-Net (DSN) has been released for complex-valued SAR images. (code: DSN_src)

any questions please contact huangzhongling15@mails.ucas.ac.cn

Environment

Pytorch 0.4.0 (also verified in Pytorch 1.4.0)
Python 3.6

SAR-Specific Models

./model/resnet18_I_nwpu_tsx.pth [1]

The SAR image pre-trained model in Reference [1].

It can be transferred to any SAR classification or detection models with ResNet-18 backbone.

Target Detection

Taking MMDetection framework for example. MMDetection is an open source object detection toolbox based on PyTorch, it provides an interface to import the pretrained model to initialize the backbone when designing the detection network.

./model/resnet18_tsx_mstar_epoch7.pth [1]

The transferred model to MSTAR 10-class target recognition task in Reference [1], achieving an overall accuracy of 99.46%.

./model/alexnet_tsx.pth [2]

The SAR-image pre-trained model in Reference [2].

./model/alexnet_tsx_mstar_iter1920.pth [2]

The transferred model to MSTAR 10-class target recognition task in Reference [2], achieving an overall accuracy of 99.34%.

./model/slc_spexy_cae_3.pth [3]

The pre-trained stacked convolutional auto-encoder model for frequency signals in Reference [3].

./model/slc_joint_deeper_3_F.pth [3]

The trained DSN model in Reference [3].

References

[1] Classification of Large-Scale High-Resolution SAR Images with Deep Transfer Learning, IEEE GRSL 2020

doi: 10.1109/LGRS.2020.2965558

[2] What, Where and How to Transfer in SAR Target Recognition Based on Deep CNNs, IEEE TGRS 2019

doi: 10.1109/TGRS.2019.2947634

[3] Deep SAR-Net: Learning Objects from Signals, submitted to ISPRS Journal of Photogrammetry and Remote Sensing 2020

doi: 10.1016/j.isprsjprs.2020.01.016

Deep SAR-Net (DSN)

Training Procedure

1. Run data_process.py

Generate the 4-D hyper-image signals of SAR images, and obtain the mean/std value for further usage.

2. Run train_cae.py

Train the stacked convolutional auto-encoder for frequency signals to obtain the cae model.

3. Run mapping_r4_r3.py

To save the computing resources, map the 4-D hyper-image signal of SAR image to 3-D tensor with the pre-trained cae model.

4. Run train_joint.py

Train the post-learning subnet and fine-tuning the image representation subnet.

Testing Procedure

1. Run mapping_r4_r3.py

Map the 4-D hyper-image signal of SAR images in test set.

2. Run test_joint.py