In order to solve the problem that stereo matching of binocular underwater image could not meet conventional epipolar constraint in the air, a semi-global algorithm for dense stereo matching of underwater image based on depth constraint was proposed. Firstly, a depth constraint was used to determine the searching area during stereo matching; based on depth constraint area, the absolute difference function and gradient calculation were extended to two dimensional directions and then fused by weighted factor. During candidate searching in the depth constraint area, the winner-takes-all was adopted to get the best line aberration under each disparity and their corresponding cost values, these cost values were regarded as data item of semi-global algorithm and an initial disparity map was obtained. Finally the sub-pixel dense disparity map was obtained by parabolic fit. Experiments were performed on underwater image to obtain dense disparity map, the results show that compare with other semi-global algorithms, the proposed algorithm could greatly accelerate underwater stereo matching and improve accuracy of matching.