SAR object classification with classical methods on MSTAR-10 class dataset.

Term project of CS554 Computer Vision course at Bilkent University.

python 3.7.10
numpy==1.19.1
opencv-python==4.5.1
Pillow==9.3.0
PyWavelets==1.1.1
scikit_image==0.17.2
scikit_learn==1.2.0
scipy==1.5.2
skimage==0.0

Public SDMS MSTAR dataset can be downloaded from this link.

• There is not an already prepared 10-class dataset of MSTAR. You need to download MSTAR Clutter and MSTAR Target Chips (T72 BMP2 BTR70 SLICY), extract 10 classes (2S1, BMP2, BDRM_2, BTR70, D7, T62, T72, ZIL131, ZSU_23_4, BTR_60) according to their test (15 degree depression angle) and train (17 degree depression angle) sets.

• After organizing the 10-class RAW files, extract JPEG, HDR and MAG files using the MSTAR PUBLIC TOOLS. You may modify the provided generate_hdr_mag.bash and use for batch processing.

For masked ground-truth data, we use SARBake¹. You may use the utility scripts in data.py to convert the masks from CSV to PNG, and you need to manually seperate 15-17 degrees of azimuth angles for each class since they are combined.

We provide the organized, pre-processed, and ready-to-use MSTAR dataset to avoid all the hassle. Each data in this dataset has 7 variations:

• RAW
• HDR
• MAG
• JPG/JPEG
• Oriented images
• Oriented + CLEANed images
• CLEANed² images

We also provide the PNG target masks of SARBake. Each data in this dataset has 3 variations:

• PNG masks
• CSV masks
• PNG of the original SAR image + mask boundaries

• We follow a pipelined structure for the MSTAR 10-class classification task, and experiment with various methods and their combinations to get the highest classification accuracy.

1. Preprocessing: Histogram equalization, CLAHE, non-local means denoising, DWT denoising, morphological operations, image rotation, image resizing
2. Feature extraction: CLEAN, Gabor filters, PCA
3. Classification: SVM

• For CLEAN, a custom point spread function $h$ is implemented for SAR ASC³, using the bandwidth $B$, center frequency $f_c$, azimuth angle $\Theta$ and center azimuth $\theta_c$ from HDR files:

$$ CLEAN(I_{noisy}(x,y), h) = I_{clean}(x,y) $$

$$ I_{clean}(x,y) = \Sigma_{n=1}^{N}A_N h(x-x_n, y-y_n) $$

$$ h(x,y) = e^{\frac{j 4 \pi f_c}{c}{(x+\theta_c y)}}{\frac{4 f_c B \Theta}{c^2}} sinc(\frac{2B}{c}x) sinc(\frac{2 f_c \Theta}{c}y) \omega(x,y) $$

$\omega$ is the -35dB Taylor window function that was originally utilized while collecting the MSTAR data.

• We observe that downscaling to 16x16 + setting all azimuth angles to 0 (image rotation) + histogram equalization + DWT denoising + 60-dim PCA features + SVM combination yields the highest accuracy (98.9%) among all tried combinations.

• Please download the two pre-processed datasets and extract to the project root folder. Make sure that the project folder directory tree looks as follows:

.
├── data.py
├── main.py
├── preprocess.py
├── dataset
│   ├── Same directory layout as SARBake
└── SARBake
    ├── TEST_15
    │   ├── 2S1
    │   ├── BMP2
    │   │   ├── SN_9563
    │   │   ├── SN_9566
    │   │   └── SN_C21
    │   ├── BRDM_2
    │   ├── BTR_60
    │   ├── BTR70
    │   │   └── SN_C71
    │   ├── D7
    │   ├── T62
    │   ├── T72
    │   │   ├── SN_132
    │   │   ├── SN_812
    │   │   └── SN_S7
    │   ├── ZIL131
    │   └── ZSU_23_4
    └── TRAIN_17
        └── Same layout as TEST_15

• python3 main.py --use_PCA=true --use_SARBake=false --case=2

• case handles which files to be read:

  • case=0 : Read hdr files
  • case=1 : Read CLEANed images
  • case=2 : Read oriented images
  • case=3 : Read oriented+CLEANed images
  • case=4 : Read vanilla MSTAR images
  • case=5 : Read png-converted csv masks of SARBake overlays

• To change the augmentation type, you may change the aug_args list, while referring to the generate_ablation_params function in main.py.