Vision-Assisted Near-Field Channel Estimation for XL-MIMO Systems

In extremely large-scale multiple-input multiple-output (XL-MIMO) near-field communications, accurate channel estimation is crucial but often requires long pilot sequences and high computational overhead, which significantly limits practical deployment. To overcome this challenge, we propose a vision-assisted near-field channel estimation method for XL-MIMO systems. The core idea is to exploit out-of-band side information obtained from vision imaging to reduce pilot length and computational complexity. To this end, we design a novel deep learning framework, termed HyperDimScaleFusionNet, which processes sequential image data to capture temporal dependencies between consecutive frames. Within this framework, we further develop a feature fusion module based on multi-scale and multi-dimensional shifted channel attention. This module strengthens the representation capability of the network by effectively extracting and integrating features across different dimensions and scales, enabling the model to fully exploit both vision and pilot information. Extensive experiments demonstrate that the proposed approach reduces the computational complexity of channel estimation by up to 90% compared with conventional methods, while achieving superior estimation accuracy. In particular, when the pilot length is short, the vision-assisted method significantly outperforms conventional sparse recovery algorithms.