The census transform in existing stereo matching algorithms has good effect in the weak texture region, but it neglects the gray level information of the image, causing unsatisfactory matching effect in repeated texture regions.  We therefore propose an improved census transform. In the initial matching cost stage, we design a similarity measure algorithm based on census transform, mutual information and gradient information. Then in the cost aggregation stage, we adopt an adaptive weight guided filtering aggregation strategy. Finally, the final disparity map is obtained by calculating and optimizing disparity. We use the proposed algorithm to test the standard images provided on the Middlebury website on the VS2015 software platform. Experimental results show that it can get an accurate disparity map and the average mismatch rate is 5.29%, which meets the requirement of 3D reconstruction.