In terms of underwater non-parallel binocular image matching not being parallel correction, and no longer satisfying the epipolar constraint in the air, a kind of epipolar constraint model that suitable for underwater is presented. According to the differences of relative position between the camera and waterproof cover, the underwater imaging ways of binocular vision is divided into parallel imaging, common reflection surface imaging and independent reflection surface imaging. And the epipolar constraint model of the last two are deduced. The camera is calibrated. The scale invariant feature transform (SIFT) algorithm can help to match stereo. And corresponding curve of the feature points on reference image is derived. According to whether the corresponding feature is on the curve, the false matching points are eliminated and the matching accuracy are improved. Experimental results show that the underwater curve constraint model can effectively eliminate false matching points, and reduce the error matching rate.