Light Field (LF) cameras can capture light rays from different angles along with their intensities, making it crucial to leverage the spatial and angular information contained in low-resolution (LR) LF images to reconstruct high-resolution (HR) LF images. Although many methods have been applied to LF image super-resolution (SR), existing approaches primarily consider disparity information in the horizontal and vertical Epipolar Plane Image (EPI) spaces of LF images, neglecting motion information in other subspaces. Consequently, these methods suffer significant performance drops when dealing with LF images exhibiting different disparities. This paper employs a generalized EPI representation that includes two additional horizontal and vertical EPI subspaces, providing motion information about objects. Furthermore, we propose a multi-scale EPI feature extraction scheme based on separable convolutions, which learns sub-pixel information of LF images from multiple dimensions. This effectively extracts and aggregates features across different disparity ranges, adapting to varying degrees of disparity changes. Based on this, we design a simple yet effective convolutional neural network, GMS-Net, and validate its effectiveness through extensive ablation experiments. Experimental results on real-world and synthetic LF datasets demonstrate that our method achieves state-of-the-art (SOTA) performance and exhibits superior robustness to disparity variations compared to other methods.