In recent years, the use of freeform optical surfaces in optical system design has experienced a significant increase, allowing systems to achieve a larger field-of-view and/or a smaller F-number. Despite these advancements, further expansion of the field-of-view or aperture size continues to pose a considerable challenge. Simultaneously, the field of computer vision has witnessed remarkable progress in deep learning, resulting in the development of numerous image recovery networks capable of converting blurred images into clear ones. In this study, we demonstrate the design of offaxis freeform imaging systems that combines geometrical optical design and image recovery network training. By using the joint optimization process, we can obtain high-quality images at advanced system specifications, which can be hardly realized by traditional freeform systems. We present a freeform three-mirror imaging system as a design example that highlights the feasibility and potential benefits of our proposed method. Zernike polynomials surface with an off-axis base conic is taken as the freeform surface type, using which the surface testing difficulty can be controlled easily and efficiently. Differential ray tracing, image simulation and recovery, and loss function establishment are demonstrated. Using the proposed method, freeform system design with increased field-of-view and entrance pupil size as well as good image recovery results can be realized. The proposed method can also be extended in the design of off-axis imaging systems consisting phase elements such as holographic optical element and metasurface.