Reconfigurable intelligent surfaces (RIS) are widely perceived as a transformative technology for 5G and beyond, enabling dynamic programming of wireless propagation channels. However, acquiring accurate channel state information (CSI) remains a major challenge in RIS-assisted wireless communication systems. Most existing research assumes the availability of full CSI, which is often impractical due to the passive nature of the RIS elements and the high-dimensional nature of the channels. To fill this gap, we introduce a two-stage framework, called the denoising super-resolution network (DSRnet), to estimate full CSI from partial representations. Then the estimated full CSI is utilized to maximize the weighted sum-rate (WSR) via phase shift prediction. DSRnet employs a hierarchical architecture consisting of a super-resolution sub-network for initial estimation, followed by a denoising sub-network enhanced with spatial attention modules for refined processing. The proposed model achieves channel estimation performance with a normalized mean square error of \(-13.77\) dB while maintaining computational efficiency, utilizing only 80, 916 parameters. When the estimated full CSI is used for phase shift prediction, it yields approximately a 15% improvement in WSR compared to methods relying on partial CSI. These findings highlight DSRnet as a practical and efficient solution for large-scale RIS deployments, effectively balancing high performance with reduced CSI overhead.