Convolutional layers typically focus on high frequency features and local details of images, yet low-frequency and non-local information are crucial for single-image super-resolution (SISR) reconstruction. To address this, we propose a lightweight multi-branch feature aggregation network (MFAN) that synergizes local and non-local information. First, the local detail enhancement block (LDEB) extracts high frequency details through multi branch convolutions and pooling operations. Second, the efficient approximation of large kernel convolutions (EALKC) employs a multi scale downsampling strategy, simulating the effect of large kernels with lightweight small kernel convolutions to expand the receptive field and capture low-frequency information. Finally, the feature aggregation block (FAB) efficiently integrates these two types of features via a cross-channel attention mechanism, where the output of EALKC serves as the query and the output of LDEB acts as the key and value. Experiments demonstrate that MFAN achieves a PSNR of 31.19 dB on the ×4 Manga109 dataset with only 463K parameters, surpassing the suboptimal model by 0.39 dB while requiring only 45% of the parameters of comparable lightweight models. The results indicate that MFAN achieves a remarkable balance between performance and lightweight design.