Model Compression Based on SAR-Optical Image Mapping Task

In recent years, due to the scarcity of resources for pairing SAR images with optical images, image mapping tasks have become an important research direction in the field of Earth observation. Conditional Generative Adversarial Networks have demonstrated their superior performance and great potential on the task of SAR-Optical image translation. However, a more complex network means a larger amount of computing and more computing resource consumption, which makes task deployment and application landing become more challenging. In this paper, we propose a compression algorithm based on the image mapping model, which can minimize the number of parameters and the computational resource costs of the model, while preserving its performance of the model most. We analyze the structure and parameter distribution of the image generator, and design a lightweight mapping module based on Depthwise Separable Convolution. In view of the sensitivity of Conditional Generative Adversarial Networks to structure, we design an automatic channel pruning algorithm based on Neural Architecture Search. This algorithm further compresses the number of parameters on the lightweight generator to speed up inference. Finally, we test on the SAR-Optical image mapping task, and the model under the compression algorithm has a better mapping effect than the model of the same scale. The algorithm achieves a better mapping effect at a lower cost of computing resources, and provides more possibilities for the deployment and development of image mapping tasks.