Object-Fidelity Remote Sensing Image Compression With Content-Weighted Bitrate Allocation and Patch-Based Local Attention

As the contradiction between high-resolution remote sensing (RS) image acquisition and limited storage space or bandwidth becomes increasingly prominent, the importance of prioritizing the compression of object regions has become evident. However, how to achieve object-fidelity RS image compression at low bitrates remains a challenging task. In this article, we use a deep neural network to develop a novel object-fidelity RS image compression (OF-RSIC) method. First, an object detection algorithm is employed to split the image into object and background regions, followed by background region smoothing through bilinear downsampling to reduce the bitrate required for background region encoding. Subsequently, a Transformer-based content-weighted attention module (CWAM) is developed for adaptive bits allocation. This module allows the model to capture global correlations among pixels and generates a more compact representation for the latent features. Additionally, a patch-based local attention module (PLAM) is proposed to reweight the local information of the entropy model, thereby improving the rate-distortion (RD) performance. Finally, to constrain the bitrate allocation between the background and object regions, a region-differentiated loss is introduced for the model training. To assess the efficacy of the proposed method, we employ an object detection algorithm to select four classes of object images from the DIOR dataset for experimentation. Comprehensive experimental results demonstrate that OF-RSIC outperforms state-of-the-art image compression algorithms in terms of object fidelity at lower bitrates.