At present, deep learning plays a crucial role in structured light 3D reconstruction. Further, in the field of fringe projection profilometry, learning 3D features from fringes and performing 3D reconstruction are being studied by many researchers. This paper combines deep learning with binocular fringe projection, uses three channels to form a single-composite-color fringe pattern for three different frequencies as the input, and predicts the numerator and denominator required to solve the wrapped phase of the object. The wrapped phase is calculated using an arctangent function. The absolute phase is obtained by unwrapping the multi-frequency heterodyne method, and the absolute phase of the left and right cameras is matched to obtain a disparity map. The parameters obtained through camera calibration can restore the 3D shape of the object, which greatly reduces the number of fringes required; accuracy close to the phase of the training set is achieved. Finally, the experimental results demonstrate the feasibility of this approach.