Author Affiliations
1College of Operational Support, Rocket Force University of Engineering, Xi'an, Shannxi 710025, China2Beijing Institute of Remote Sensing Equipment, Beijing 100854, Chinashow less
Fig. 1. Structure of the pyramidal multi-scale GAN
Fig. 2. Structure of generator
Fig. 3. Inception block
Fig. 4. Linear convolutional layer and 1×1 convolutional layer. (a) Linear convolution layer; (b) 1×1 convolution layer
Fig. 5. Residual dense block
Fig. 6. Structure of the discriminator
Fig. 7. Image generated from single image. (a) Image1 of small ship; (b) image2 of small ship; (c) image with noise in background; (d) image of large ship
Fig. 8. Images generated by different networks. (a) Images used for training; (b) images generated by the original network; (c) images generated by the improved network
Fig. 9. Error detection of training on different data sets. (a) Correct test results; (b) SSDD data set; (c) SSDD data set + generated sample data set
Fig. 10. False alarms of training on different data sets. (a) Correct test results; (b) SSDD data set; (c) SSDD data set + generated sample data set
Fig. 11. Missed detection of training on different data sets. (a) Correct test results; (b) SSDD data set; (c) SSDD data set + generated sample data set
Fig. 12. Undetected result after adding the generated data set. (a) Correct test results; (b) SSDD data set; (c) SSDD data set + generated sample data set
Block | Operation | Convolution kernel | Input channel | Output channel |
---|
Head | Conv_block | 3×3 | 3 | 192 | Inceptionblock | Block 1 | Conv_block | 1×1 | 192 | 64 | Block 2 | Conv_block | 1×1 | 192 | 96 | Conv_block | 3×3 | 92 | 128 | Block 3 | Conv_block | 1×1 | 192 | 16 | Conv_block | 3×3 | 16 | 32 | Conv_block | 3×3 | 32 | 32 | Block 4 | Max Pooling | Conv_block | 1×1 | 192 | 32 | Connection part | Conv_block | 3×3 | 256 | C | | Residual dense block | Conv | 3×3 | C | C | Leaky ReLU | Conv | 3×3 | 256 | C | Leaky ReLU | Conv | 3×3 | 384 | C | Leaky ReLU | Conv | 3×3 | 512 | C | Leaky ReLU | Conv | 3×3 | 640 | C | Leaky ReLU | tail | Conv_block | 3×3 | C | C | Conv_block | 3×3 | C | C | Conv | 3×3 | C | 3 | | tanh |
|
Table 1. Parameters of the generator
Dataset | AP |
---|
SSD | Tiny-YOLO |
---|
SSDD | 0.8097 | 0.707 | SSDD_method1_20 | 0.7275 | 0.641 | SSDD+ SSDD_method1_20 | 0.8178 | 0.726 | SSDD+ SSDD_method1_40 | 0.8101 | 0.716 | SSDD+ SSDD_method2_20 | 0.8106 | 0.696 | SSDD+ SSDD_method2_40 | 0.7907 | 0.672 | SSDD+ SSDD_method2_1160 | 0.8213 | 0.702 |
|
Table 2. AP of different methods to generate images