Abstract
Traditional Generative Adversarial Networks (GANs) often struggle with uncontrollable generation processes and a lack of diversity in synthesized remote sensing images. This paper proposes a novel generative model based on background conditions and visual attention mechanisms. By employing a U-Net generator and a color-coded conditional mask, the model focuses on learning specific target features while ignoring non-essential information. Experimental results indicate that the proposed method achieves a Peak Signal-to-Noise Ratio (PSNR) of 18.512 and a Structural Similarity Index (SSIM) of 88.47%, significantly outperforming standard autoencoder-based models.
1. Introduction
Training robust deep learning models for remote sensing target detection requires vast amounts of diverse data. However, acquiring high-resolution satellite imagery is often expensive and logistically challenging. While traditional data augmentation (e.g., rotation, cropping) increases sample size, it does not introduce new structural diversity. Generative models like GANs offer a path forward, but standard architectures lack the precision required to place specific targets in precise locations. This study addresses these limitations by introducing a Conditional GAN (CGAN) framework tailored for remote sensing applications.
2. Proposed Methodology
The proposed method transitions from "black box" generation to a controllable process using spatial and categorical constraints.
2.1 Conditional Mask and Attention Mechanism
To provide precise control, the researchers implemented a color-coded conditional mask. Different RGB values in the mask correspond to specific target categories, such as standard ships or aircraft carriers. This mask serves as a visual attention mechanism, directing the network to synthesize targets at specific coordinates while preserving the surrounding environment.
2.2 Network Architecture
The generator utilizes a U-Net structure with eight layers of down-sampling and eight layers of up-sampling. Skip connections are integrated to pass low-level background information directly across the network, ensuring that the synthesized targets blend realistically with the sea texture without distorting the original background.
2.3 Loss Function and Discriminator
The model employs an "Area Discriminator" that evaluates the image in blocks rather than as a single scalar. This forces the generator to maintain high fidelity in local details. The training is guided by a hybrid loss function combining Least Squares Loss (LSGAN) for stability and L1 regularization to minimize the pixel-wise difference between the generated target and the background.
3. Experimental Results
The model was evaluated using a dataset of 420 high-resolution sea surface image pairs ($256 \times 256$ pixels). The proposed CGAN method was compared against two baseline autoencoder models with varying residual blocks.
| Model | PSNR | SSIM |
|---|---|---|
| Proposed CGAN Method | 18.512 | 88.47% |
| Autoencoder (9 residual blocks) | 17.596 | 81.65% |
| Autoencoder (6 residual blocks) | 17.089 | 76.31% |
The results demonstrate that the CGAN approach produces sharper images with higher structural consistency compared to the baselines.
4. Conclusion
This research presents an effective method for generating controllable and diverse remote sensing target images. By combining U-Net architectures with conditional masks, the model successfully addresses the data scarcity problem in satellite imagery analysis. Future work will explore applying this method to more complex land-based targets and larger-scale datasets.
Summary updated