Grounded in the two-plane 4D light field (LF) model, LF cameras capture both spatial and angular information of 3D scenes at the expense of spatial resolution. To improve the spatial resolution of LF images, in this paper, a LF image super-resolution network based on iterative feature separation and fusion is proposed. A pyramid blueprint convolution block is designed based on the intra-kernel correlation, which is used to accomplish feature separation and fusion. In the initial separation stage, two groups of PBCB cascaded residual block are employed to extract spatial characteristics from each sub-aperture image. During the deep iterative fusion and separation stage, each iterative unit first fuses the rearranged features via the feature separation and interaction module, then re-separates the fused features using the PBCB. For data reconstruction, the nested residual distillation block maps the final separated features into high-resolution image details, which are further processed by the upsampling module to produce the final super-resolution image. Comparative experiments confirm that the designed method surpasses state-of-the-art LF super-resolution methods, achieving an average peak signal-to-noise ratio of 39.16 dB on the 2 \(\times \) super-resolution task across five public datasets.