Deformable image registration can align two images acquired at different times or under different conditions, and thus has important theoretical research and clinical application value. However, existing learning-based registration methods are typically narrowly restricted to small displacement deformations. So to overcome this difficulty, we propose a recursive cascaded parallel network for image registration (RCPN). RCPN optimizes the image registration performance in a coarse-to-fine manner by multiple cascades to incrementally enhance the registration of target image pairs. In addition, the cascaded subnetwork always maintains a high-resolution representation. It connects multi-resolution streams in parallel, thus strengthening the feature representation capability in a finite perceptual field, which allows the registration network focus more on the small displacement deformations beyond the large displacement deformations of the image in space and enhances the registration capability. The experiments based on standard MRI images of the brain indicate that RCPN has significantly improved the registration performance compared with the benchmark method, and the effectiveness of this method can be effectively verified by combining other comparative results.