SAR Target Classification Based on full convolution network

1. MSTAR dataset expansion

• In the original SAR image (128) × 128) random clipping 88 × 88 slices, each slice can contain the target area, and the number of samples can reach (128-88 + 1) after random sampling × (128-88 + 1) = 1681 times, each SAR image of each category samples 10 slices, so as to improve the number of training samples.

2. Full convolution network model (AConvNets)

• The network is completely realized by the convolution layer, the full connection layer is removed, the convolution output of the last layer directly outputs the probability of each category by Softmax, and the loss function is still the loss of classification cross entropy. The specific network model is as follows:

• Among them, the convolution layer uses the BN layer to prevent over fitting and accelerate training. The input training set dimension: [b, 88, 88, 1] - > output: [b, 1, 1, 10]. The classification cross entropy loss is still used for training. The optimizer Adam has a learning rate of 0.0001.

3. Classification results under SOC conditions

• The training set and test set are as follows:

  category	quantity
  	Training set 17 °	Test set 15 °
  2S1	299	274
  BMP2	233	196
  BRDM2	298	274
  BTR60	256	195
  BTR70	233	196
  D7	299	274
  T62	299	273
  T72	232	196
  ZIL131	299	274
  ZSU23/4	299	274
  total	2747	2426

• Each SAR image in the training set is cropped with 5 random slices and 1 slice in the center, all of which are 88 × 88 pixels, a total of 2747 training sets × 6 = 16482 (sheets). Test set center clipping 88 × 88 slices.

• Training: epoch = 50, Adam optimizer, learning_rate = 0.0001, the Accuracy and Loss curves of training set and test set are displayed in the tensorboard:

  

• Calculate the confusion matrix of the test set as follows:

   	 2S1 BMP2 BRDM B60 B70 D7 T62 T72 ZIL ZSU	
  2S1[[263   0   0   0   0   0  11   0   0   0]	95.99%
  BMP2[  0 132   0  18   3   0   0  42   0   1]	67.35%
  BRDM[  0   0 272   0   0   0   0   0   0   2]	99.27%
  B60 [  0   0   0 192   0   0   0   0   1   2]	98.46%
  B70 [  0   0   0  13 181   0   0   2   0   0]	92.35%
  D7 	[  2   0   0   0   0 272   0   0   0   0]	99.27%
  T62 [  0   0   0   0   0   0 273   0   0   0]	100%
  T72 [  0   0   0   1   0   0   0 195   0   0]	99.49%
  ZIL [  0   0   0   0   0   4   0   0 270   0]	98.54%
  ZSU [  0   0   0   0   0   0   0   0   0 274]]	100%
  Total: 95.07% 
  The Precision is :  0.9574354589252246
  The Recall is :  0.9507109053615832
  The Accuracy is :  0.9579554822753503
  The F1 is :  0.9493873742571258
  The F_beta is :  0.949223632328047
  The Auc Score is :  0.9996171502953114
  

4. Classification results under eoc1

• EOC1 is the training set and test set under the condition of large elevation change, as follows:

  category	quantity	quantity
  	Training set 17 °	Test set 30 °
  2S1	299	288
  BRDM2	298	287
  T72	299	288
  ZSU234	299	288
  total	1195	1151

• Each SAR image in the training set is cropped with 5 random slices and 1 slice in the center, all of which are 88 × 88 pixels, 1195 training sets in total × 6 = 7170 (sheets). Test set center clipping 88 × 88 slices.

• Training: epoch200, Adam optimizer, learning_rate = 0.0001, the Accuracy and Loss curves of training set and test set are displayed in the tensorboard:
  [external chain picture transfer failed. The source station may have anti-theft chain mechanism. It is recommended to save the picture and upload it directly (img)-
  

• Calculate the confusion matrix of the test set as follows:

       2S1  BRDM  T72  ZSU	
  2S1   288   0    0    0		100%
  BRDM   0   287   0    0		100%
  T72    0    0   287   1		99.65%
  ZSU   59    0    0   229	79.51%
  Total: 94.79% 
  
  The CNN Precision is :  0.9564058388673098
  The CNN Recall is :  0.9479166666666666
  The CNN Accuracy is :  0.947871416159861
  The CNN F1 is :  0.9473793419770824
  The CNN F_beta is :  0.9465925313694183
  

5. Classification results under eoc2 conditions

• EOC2 is the change of vehicle appearance configuration, and its training set is as follows:

  category	Quantity (pitch angle 17 °)
  BMP2(9563)	233
  BRDM2	298
  BTR70(C71)	233
  T72(132)	232
  total	995

• Test set:

  category	model	Quantity (pitch angle 17 ° & 15 °)
  T72	S7	419
  T72	A32	572
  T72	A62	573
  T72	A63	573
  T72	A64	573
  total		2710

• Each SAR image in the training set is cropped with 10 random slices and 1 slice in the center, all of which are 88 × 88 pixels, 995 training sets in total × 11 = 10945 (sheets). Test set center clipping 88 × 88 slices.

• Training: epoch50, Adam optimizer, learning_rate = 0.0001, the Accuracy and Loss curves of training set and test set are displayed in the tensorboard:
  

• Calculate the confusion matrix of the test set as follows:

  	BMP2  BRDM2  BTR70   T72  
  S7	 8		0	   0     411	98.09%
  A32	 0      0      0     572	100%
  A62  1      0      0     572	99.83%
  A63  1      0      0     572	99.83%
  A64  4      0      0     569	96.14%
  Total: 98.78%
  

6. Classification results under eoc3

• EOC3 is for different models and variants of the same target. The training set is consistent with EOC2 training set. The test set is as follows:

  category	model	Quantity (pitch angle 17 ° & 15 °)
  BMP2	9566	428
  BMP2	C21	429
  T72	812	426
  T72	A04	573
  T72	A05	573
  T72	A07	573
  T72	A10	567
  total		3569

• Here, the EOC2 trained model is directly used for prediction. The confusion matrix of the test set is as follows:

  		   BMP2   BRDM2    BTR70   T72  
  BMP9566	   359		0		17	    52		83.88%
  BMPC21	   372		0		1		2		86.71%
  T72_812		2		0		1	   423		99.30%
  T72_A04		3		0		0	   570		99.48%
  T72_A05		1		0		0	   572		99.83%
  T72_A07		0		0		0	   573		100%
  T72_A10		20		0		0	   547		96.47%		
  Total: 95.74%