Remote Sensing Image Compression Based on High-Frequency and Low-Frequency Components
Shao Xiang,Qiaokang Liang
DOI: https://doi.org/10.1109/tgrs.2023.3349306
IF: 8.2
2024-02-02
IEEE Transactions on Geoscience and Remote Sensing
Abstract:With the increasing volume of high-resolution satellite images, image compression technology has become a research hotspot in the field of remote sensing image processing; however, the existing remote sensing image compression methods, such as JPEG2000, fail to ensure high-ratio and high-fidelity compression. To address this issue, we use a deep neural network to build a learned image compression model named HL-RSCompNet, which is specifically designed for remote sensing images. This model considers both high-frequency and low-frequency features in remote sensing images. We use discrete wavelet transformation (DWT) to divide the image features into two components: the high-frequency feature component and the low-frequency feature component. In addition, we introduce a frequency domain encoding-decoding module with the goal of bolstering the model's capacity to represent both high-frequency and low-frequency features effectively. This approach allows the model to preserve more high-frequency information, thereby enhancing the overall compression performance of the learned image compression model. Extensive experimental and validation works are performed on four high-resolution remote sensing image datasets. The results indicate that our method outperforms existing traditional compression methods like JPEG2000 and even surpasses the performance of state-of-the-art learned image compression models. Our project is available at https://github.com/shao15xiang/HL-RSCompNet.
imaging science & photographic technology,remote sensing,engineering, electrical & electronic,geochemistry & geophysics