3.3.1. Quantitative Analysis of the Image Quality
To evaluate the quality of the generated SAR deception jamming templates with shadows, its targets, shadows, and speckle noise were assessed.
The equivalent number of looks (ENL) was used to measure the relative strength of the speckle noise in the deceptive jamming templates [
39,
40,
41,
42]. The ENL depicted the smoothness of the images and reflected the speckle character, since speckle noise appears grainy and may add false edges to an originally smooth region [
43]. A lower ENL value indicated a greater presence of speckle noise. The ENL was calculated by the following.
where
represents the mean value of the SAR image, and
represents the standard deviation of the SAR image. The ENL was calculated for the background regions in each image.
The correlation coefficient (CC) describes the statistical correlation between two images. The more similar the two images, the greater the correlation coefficient [
44]. For the original image
and the generated image
, with the size of
, the average gradient index can be defined as the following.
where
and
are the mean values of
and
, respectively.
The gradient-based structural similarity (GSSIM) combines the information of the luminance comparison
, contrast comparison
, and gradient-based structure comparison
in the image, which is shown in the equation below [
45].
where
where
and
are the mean values of
and
, respectively, which reflect the luminance comparison information.
and
are the standard deviation of
and
, respectively, which reflect the contrast comparison information.
and
represent the gradient values of the pixels in row
and column
of
and
, respectively.
,
, and
are the small constants that prevent the denominator from being zero. The parameters of
,
, and
are greater than zero. In this paper,
,
, and
. The higher the GSSIM value, the more similar
and
will be [
46].
The average gradient (AG) can reflect the presentation abilities of the image details and textures, which is always used to assess the image sharpness [
47]. For the evaluated image
with the size of
, the average gradient index can be defined as the following.
where
are the coordinates of the image;
and
denote the horizontal and vertical gradient values, respectively. The larger the average gradient value, the richer information contained in the image, and the better the fused result.
The mean squared difference (MSD) is frequently used to measure the difference between the values predicted by a model, and it is utilized to evaluate the fluctuation of the gray value of the image and the degree of focus of the image [
48]. For a given
deceptive jamming template
, the MSD can be defined as the following.
where
is the average gray value of the given deceptive jamming template
. A larger MSD corresponds to a clearer image.
The evaluation results shown in
Figure 8 regarding the above four evaluation indicators are presented in
Table 1.
Table 1 shows that the original image shown in
Figure 8a had an ENL value of 2.5604. The generated SAR deception jamming templates with shadows had ENL values of 1.9764, 1.9090, and 1.9092, with an average value of 1.9315. Therefore, the difference between the ENL values of the generated templates and the original image was small, indicating a high similarity between the images of 75.44% (the ratio of the ENL between the two images). This result suggested that the generated SAR deception templates had a high authenticity.
The AG value of the original image shown in
Figure 8a was 2.7145. The AG values of the generated SAR deception jamming templates with shadows were 2.7114, 2.6935, and 2.7235, respectively, with an average value of 2.7095. Therefore, the difference between the AG values of the generated template and the original image was small, indicating a high similarity of 97.41% (the ratio of the AG between the two images). This indicated that the generated SAR deception template had a high texture and structural similarity with the original image.
The MSD value of the original image shown in
Figure 8a was 0.1540. The MSD values of the generated SAR deception jamming templates with shadows were 0.1580, 0.1580, and 0.1583, respectively, with an average value of 0.1581. Therefore, the difference between the MSD values of the generated template and the original image was small, indicating a high similarity of 97.41% (the ratio of the MSD between the two images). This indicated that the generated SAR deception template had a high similarity of grayscale and focusing characteristics with the original image.
For the original image and the generated SAR deception templates, the GSSIM and CC values of the target and shadow regions were calculated. The GSSIM values of the target and shadow regions of the original image and the generated templates were 0.9518, 0.9516, and 0.9546, with an average value of 0.9527, which indicated a high similarity of 95.27% between the generated and original images. The CC values of the target and shadow regions of the original image and the generated templates were 0.9964, 0.9963, and 0.9967, with an average value of 0.9965, which indicated a high similarity of 99.65% between the generated and original images. Therefore, the proposed scheme could generate SAR deception jamming templates with shadows that could exhibit a high similarity to the original image and have a high authenticity.
The following was an experiment on the truck (ZIL-131) and a calculation of the evaluation indicators. The generated images were shown in
Figure 9.
Table 2 showed that the similarities of the ENL value, the AG value, the MSD value, the GSSIM value, and CC value between the generated templates and the original image were 0.8333, 0.9627, 0.9833, 0.8626, and 0.9213, respectively, which indicated that the generated SAR deception template had a high texture and structural similarity with the original image and a high authenticity.
Table 2.
Evaluation index calculation results.
Table 2.
Evaluation index calculation results.
Image | ENL | AG | MSD | GSSIM | CC |
---|
Figure 10a (the original image) | 3.0648 | 4.3137 | 0.1200 | — | — |
Figure 10b (the first sample) | 2.8589 | 4.1149 | 0.1167 | 0.8562 | 0.9060 |
Figure 10c (the second sample) | 2.4291 | 4.3661 | 0.1206 | 0.9117 | 0.9543 |
Figure 10d (the third sample) | 2.3734 | 3.9767 | 0.1166 | 0.8200 | 0.9036 |
Average of the samples | 2.5538 | 4.1526 | 0.1180 | 0.8626 | 0.9213 |
Similarity | 0.8333 | 0.9627 | 0.9833 | 0.8626 | 0.9213 |
3.3.2. Comparison with the SinGAN Scheme
This experiment was performed on shadowed T72 tank samples from the MSTAR dataset. Since the SinGAN is currently one of the schemes that could perform sample augmentation on the shadowed SAR deception jamming template, the SinGAN scheme was employed for the image sample augmentation of the SAR deception templates with shadows, resulting in a dataset of 48 T72 tank SAR deception templates, as shown in
Figure 11. The comparison between the real SAR deception jamming template with shadows of the T72 tank and the SAR deception jamming template with shadows of the T72 tank generated by the proposed network is presented in
Figure 11. Due to the poor quality of the generated images, three of the better images were selected for comparison.
The evaluation results shown in
Figure 12 regarding the above four evaluation indicators are presented in
Table 3.
The results in
Table 3 showed that the ENL value of the original image shown in
Figure 12a was 2.5604, and the ENL values of the generated SAR deception jamming templates with shadows were 1.6677, 1.7150, and 0.5637, with an average value of 1.3155. The average ENL of the generated templates showed a significant difference from the original image’s ENL value, indicating a low similarity of only 51.38%. This result suggested that the generated templates exhibited weaker speckle noise than the proposed approach, resulting in a lower authenticity.
The AG value of the original image shown in
Figure 12a was 2.7145. The AG values of the generated SAR deception jamming templates with shadows were 2.3372, 2.1778, and 2.4288, respectively, with an average value of 2.3146. Therefore, the difference between the AG values of the generated template and the original image was significant, indicating a low similarity of 85.27% between the images. This indicated that the generated SAR deception template had a low texture and structural similarity with the original image.
The MSD value of the original image shown in
Figure 12a was 0.1540. The MSD values of the generated SAR deception jamming templates with shadows were 0.1404, 0.1429, and 0.1766, respectively, with an average value of 0.1533, indicating a high similarity of 99.55% between the images. This indicated that the generated SAR deception template had a high similarity of grayscale and focusing characteristics with the original image.
For the original image and the generated SAR deception templates, the GSSIM and CC values of the target and shadow regions were calculated. The GSSIM values of the target and shadow regions of the original image and the generated templates were 0.5998, 0.5710, and 0.5514, with an average value of 0.5741, which indicated a low similarity of 57.41% between the generated and original images. The CC values of the target and shadow regions of the original image and the generated templates were 0.6492, 0.6353, and 0.8903, with an average value of 0.7249, which indicated a low similarity of 72.49% between the generated and original images. Therefore, the SinGAN scheme could generate SAR deception jamming templates with shadows that could exhibit a low similarity to the original image and have a low authenticity.
The following was an experiment on the truck (ZIL-131) and a calculation of the evaluation indicators. The generated images were shown in
Figure 13.
Table 4 shows that the MSD value similarity was high, which indicated that the generated SAR deception template had a high similarity of grayscale and focusing characteristics with the original image. However, the similarities of the ENL value, the AG value, the GSSIM value, and CC value between the generated templates and the original image were 0.2181, 0.5816, 0.3677, and 0.1545, respectively, which indicated that the generated SAR deception template had a low texture and structural similarity with the original image and low authenticity.
Table 4.
Evaluation index calculation results.
Table 4.
Evaluation index calculation results.
Image | ENL | AG | MSD | GSSIM | CC |
---|
Figure 14a (the original image) | 3.0648 | 4.3137 | 0.1200 | — | — |
Figure 14b (the first sample) | 0.5175 | 2.4920 | 0.1198 | 0.3569 | 0.2734 |
Figure 14c (the second sample) | 0.6091 | 2.3711 | 0.1206 | 0.4121 | 0.1214 |
Figure 14d (the third sample) | 0.8786 | 2.6636 | 0.1300 | 0.3341 | 0.0686 |
Average of the samples | 0.6684 | 2.5089 | 0.1235 | 0.3677 | 0.1545 |
Similarity | 0.2181 | 0.5816 | 0.9717 | 0.3677 | 0.1545 |
3.3.3. Supplementary Experiments
The following two sets of supplementary experiments were performed using SAR tank (T72) images.
Firstly, we removed the spatial attention mechanism block from the generator without changing the other conditions, and the generated images are shown in
Figure 15.
Table 5 shows that the similarities of the AG value, the MSD value, the GSSIM value, and CC value between the generated templates and the original image were 0.8798, 0.8161, 0.5574, and 0.2623, respectively, which indicated that the generated SAR deception template had a low texture and structural similarity with the original image. However, the original image shown in
Figure 16a had an ENL value of 2.5604. The generated SAR deception jamming templates with shadows had an average ENL value of 2.8903, indicating a high similarity between the generated templates and the original image of 88.59%. It can be seen that when the SAM block was removed, the targets in the images had different appearance shapes, and they could not be identified as the shape of a tank. The low similarity between the tank and its shadow in the generated deception jamming template and the original image indicated that the quality of the generated deception jamming template was poor. This result suggested that the SAM block improved the network’s learning ability for the targets and their shadows.
Secondly, we replaced the mixed noise input into the generator with Gaussian noise and performed the experiment without changing the other conditions. The generated images are shown in
Figure 17.
Table 6 shows that the similarities of the AG value, the MSD value, the GSSIM value, and CC value between the generated templates and the original image were 0.9818, 0.9853, 0.8651, and 0.9811, respectively, which indicated that the generated SAR deception template had a high texture and structural similarity with the original image. However, the original image shown in
Figure 18a had an ENL value of 2.5604. The generated SAR deception jamming templates with shadows had an average ENL value of 1.6314, indicating a low similarity between the generated templates and the original image of 63.71%. This result suggested that the generated SAR deception templates had a low authenticity. From this, it can be seen that the speckle noise improved the authenticity of the generated deception jamming templates.