|
Remote sensing image radiometric resolution transformation plays a fundamental role in data storage, transmission efficiency, processing speed, image visualization, data simplification, and analysis. As a crucial preprocessing step, radiometric resolution transformation is commonly based on gray-level stretching methods. However, these methods are often time-consuming due to the complexities involved in method selection and parameter adjustment. This paper proposes a method based on Generative Adversarial Networks (GANs), inspired by histogram matching, to achieve radiometric resolution transformation of remote sensing images. Given the significant dynamic range changes in radiometric resolution transformation, establishing a direct mapping relationship is challenging. Therefore, a Discrete Histogram Approximation Loss is designed to compute the similarity between High Radiometric Resolution Images(HRRI) and Standard-Resolution Radiometric Images (SRRI), strengthening the mapping correlation and optimizing the visual features of the generated images. Remote sensing images contain a variety of ground objects and complex detail textures; hence, Structure Correction Loss is used to constrain the structural features, ensuring the clarity and accuracy of the resultant images. This study was validated using a self-made dataset, and comparative experiments with other methods were conducted. The results demonstrate that the proposed method achieves a peak signal-to-noise ratio of 29.20dB and a structural similarity index of 86.8%, significantly outperforming other methods. These findings confirm the feasibility of the proposed approach, with the generated images exhibiting high color and structural accuracy and realism.
|
