Master-slave camera systems, consisting of wide-field and pan-tilt-zoom (PTZ) camera, are widely applied in surveillance. They can monitoring the wide scene, and the high-resolution details of interesting target can be captured by PTZ camera. In order to achieve this function, the accurate cooperative calibration for these system is a prerequisite. However, the nonlinear changing PTZ parameters (e.g. intrinsic and extrinsic) with pan, tilt and zoom lead to inaccurate calibration by existing methods. What's more, the process of traditional step-by-step calibration method makes accumulative error. In this paper, we provide a new end-to-end deep neural network for cooperative calibration. This network establishes a mapping relationship between pix coordinate in wide-field camera and control parameters of the PTZ camera. By this model, the control parameters of the PTZ camera can be acquired without any complex camera calibration operation. Experiments show that the proposed neural network has little calibration errors as compared to the ground truth.