This paper introduces a deep learning-based approach for reconstructing multi-contrast cardiac magnetic resonance images (CMRI) from highly undersampled k-space data. Traditional methods, such as parallel imaging and compressed sensing, often suffer from slow reconstruction times and limited acceleration rates. To address these challenges, we propose an advanced unrolled network architecture that separately optimizes low-frequency and high-frequency components, leveraging diverse information from k-space, the image domain, and temporal sequences. This method demonstrates robust performance across various imaging sequences and anatomical views, including long-axis, short-axis, and aortic views. The proposed approach, evaluated on the CMRxRecon challenge dataset, effectively improves image quality while maintaining clinical interpretability. Our results show improvements in PSNR, SSIM metrics compared to traditional methods, highlighting the potential of this approach to enhance CMRI reconstruction and expand its clinical applicability.