As science and technology develop, there is a non-negligible contradiction between the cost of the shooting system and the requirements for image quality in the machine vision system. Light field cameras are of great significance for 3D reconstruction, virtual reality and other fields, where there is a high demand for high quality images. However, high-quality component images imply a high cost. Super-resolution, as an image processing technology, provides the possibility of balancing high quality with low cost. Deep learning-based super-resolution methods outperform traditional methods in image quality enhancement but require plenty of high-resolution images for training. Therefore, we explored how to use low-resolution light field images and a small number of high-resolution images in the same scene to generate corresponding high-resolution images. This research utilized a network based on an attention mechanism to achieve reference-based super-resolution and improve light field image clarity. Experiments were conducted on a public light field image dataset, and the results show that the method greatly improves the resolution of the original image.