Stereo matching is the key to binocular vision measurement, which extracts depth information from left and right images captured by binocular cameras to achieve three-dimensional measurement of the target. Reconstructing the three-dimensional morphology of sample surface by a binocular vision system can facilitate the quantification of product surface quality information, characterize defects during the manufacturing process of the product, and assist in analyzing the distribution patterns of product defects. However, due to factors such as an unstable physical environment, the geometric shape of the surface being measured, and the precision of the acquisition equipment, existing stereo matching algorithms are difficult to balance accuracy and real-time performance simultaneously, which can affect the efficiency of industrial testing. How to improve the accuracy of stereo matching of binocular images and enhance the measurement accuracy of binocular vision is still the main problem facing this research field. For these reasons, the stereo matching model is established based on binocular visual imaging system, and a dynamic weight cost aggregation stereo matching algorithm based on adaptive windows is proposed in this manuscript.Firstly, traditional local matching algorithms usually use a single weight for cost aggregation under different aggregation windows, while ignoring the differences between pixels in different regions, which can easily lead to unstable stereo matching accuracy based on binocular vision measurement. a cost aggregation adaptive cross window is constructed based on the gradient information representation model as a constraint to adapt to the different requirements of window size in weak texture regions and disparity discontinuous regions in this paper. The algorithm proposed in the paper can achieve a large adaptive window in weak texture regions and can limit its arm length extension in texture rich regions.Secondly, analyze the pixel features of discontinuous disparity regions and weak texture regions, a cost aggregation model is established based on pixel distance and color difference dual threshold weights to calculate the dynamic weight influence factors of each window, which can achieve the distribution of cost weights for different windows. In terms of cost aggregation performance testing, the comparative experiment with the AD-Census algorithm shows that the average mismatch rate of the proposed algorithm in the paper is 4.21%, and its overall matching accuracy has a significant advantage.Thirdly, in order to recover the information of invalid pixels, a local neighborhood of invalid pixels is constructed based on the cross intersection method, and then the occluded points and mismatched points are interpolated and filled separately to obtain a denser disparity image. Additionally, the disparity image area is segmented based on the linear iterative clustering method. By utilizing the mean and variance information of local regions, singular disparity values are removed, and reliable pixel disparity values are searched for to fill in, thereby improving the overall matching accuracy of the disparity map.Finally, the experimental results show that in the testing of the Middlebury dataset, the proposed algorithm has an average mismatch rate of 4.11% for non-occluded areas and 5.65% for all areas, respectively, which is better than traditional matching algorithms. Based on the algorithm proposed in this paper, a binocular system platform was constructed for experiments, and the measurement results of 3D printed samples were compared with those obtained by the triangular laser method. For 4 groups of measured samples, the average relative error of global length measurement is less than 1.2%, and the average relative error of global height measurement is less than 2.7%. The experimental results verify the effectiveness and reliability of the algorithm proposed in this paper.