Abstract: To address detail loss in complex feature regions of neural network-based satellite image compression, a novel algorithm integrating region-aware mechanisms with variational autoencoder (VAE) quantization was proposed. The method prioritized key-region quality while maintaining overall compression ratios and enabling covert data embedding. Firstly, critical regions were identified by YOLO, then mapped images to latent space via ResNet VAE. During quantization, key-region features were hidden within background bitstreams before entropy encoding. Differential compression strategies were applied: aggressive compression for backgrounds and mild compression for key regions. Experiments demonstrated superior performance over conventional and deep learning benchmarks. At 25×compression ratio, the method achieved 35.27dB PSNR (full image), surpassing state-of-the-art techniques by 3.5dB. Key regions attained 41.15dB PSNR and 0.992 SSIM, a 7.55dB improvement over baseline algorithms, effectively preserving fine details. The algorithm enhances compression fidelity without increasing bitstreams, while enabling secure data hiding. It proves particularly effective for high-resolution satellite imagery and sensitive area protection, offering a balanced solution for efficient storage and secure transmission.

Key words: satellite image compression, region perception, variational autoencoder network, data hiding, differential compression