Feature representation learning is the task of learning appropriate low-dimensional representations from raw high-dimensional data. Graph learning is a feasible method for this task. Due to the shallow graph learning approaches can not capture deep structure, deep graph learning approaches like Graph Neural Networks (GNNs) are more suitable approaches for mining the deep structure within data. However, GNNs cannot be directly applied to image data. Existing approaches applying GNNs to image data have issues with high computational complexity, large memory consumption, and difficulties in data annotation. Therefore, we proposed a novel learning framework to deal with this problem. We first construct the affinity graph based on the k-nearest neighbor method, then utilize graph curvature to revise the weight relationships between graph nodes, and finally integrate revised weight information into GNN to learn the representation of nodes. Experiments on public datasets demonstrate that our method outperforms both the methods based on shallow graph learning and the methods of direct application of GNNs to image data, validating the effectiveness of our approach.