Underwater processing and manufacturing plays an important role in aviation, shipbuilding and other fields. In situ online 3-D detection of manufacturing process has become an urgent demand for underwater manufacturing quality assurance. Fringe projection profilometry (FPP), as one of the classic optical 3-D measurement technologies, holds the advantages of non-contact, high speed and high accuracy. However, in turbid water, due to the absorption and scattering of light, the fringe light intensity captured by the camera is attenuated, the contrast is reduced, the image details are blurred, and a lot of noise is introduced, resulting in poor fringe image quality. The phase calculated by the low-quality fringes has the non-negligible phase error, resulting in the decrease of the 3-D measurement accuracy. In order to reduce the influence of underwater absorption and scattering, an end-to-end fringe image enhancement algorithm based on deep learning is proposed. The fringe pattern enhancement convolutional neural network (FPENet) is used to convert the low contrast and high noise fringes into high contrast and low noise fringes to obtain more accurate phase results. FPENet can effectively improve fringe quality and reduce phase error for water with different turbidity. Especially in high turbidity water, the phase error can be reduced by about 50%, significantly improving the measurement accuracy of underwater FPP, which is of great significance for improving the applicability of FPP in complex scenes.