DCER：双阶段压缩与基于能量的图像重建

This paper introduces DCER, a novel image compression and reconstruction framework designed to address the inherent trade-off between compression ratio and reconstruction quality in existing methods. DCER's core innovation lies in its dual-stage compression strategy. Initially, a deep learning model is employed for efficient feature extraction and preliminary compression of the original image, yielding a compact latent representation. Subsequently, this latent representation undergoes further refinement and compression via an energy-based model, which is adept at capturing more intricate structures and textural information within the image. For the reconstruction phase, DCER utilizes an innovative energy-based mechanism that effectively recovers high-quality images from the compressed representation. Unlike conventional reconstruction approaches relying on mean squared error or perceptual losses, DCER's energy model guides the reconstruction process by learning the low-dimensional manifold of the data distribution. This approach significantly enhances the visual quality and detail preservation of reconstructed images while maintaining high compression rates. Experimental results demonstrate that DCER consistently outperforms state-of-the-art image compression and reconstruction techniques across various standard datasets, exhibiting particularly superior performance at low bitrates. This breakthrough opens new avenues for research in image compression. Furthermore, DCER's modular design facilitates its integration with other advanced technologies, promising broad applicability.