To solve the problem that the existing stereo-matching methods based on 2D and 3D convolutions have poor matching effects in the edge region, a multiscale cost-aggregation stereo-matching method was proposed by introducing an attention mechanism. Firstly, the parameter-free attention mechanism was embedded into the atrous spatial pyramid pooling (ASPP) module for multi-scale feature extraction of left and right images. Then, the feature maps obtained using a convolutional network were used to construct the costums, and 3D convolution was used to extract the left and right feature maps and the information between different parallax levels for feature fusion. Finally, the feature-fused costume was mapped back to the resolution of the original image by using transposed convolution to obtain the optimal parallax value. Tests on KITTI 2015 and SceneFlow datasets showed that the proposed method had the optimal average endpoint error (EPE) and the proportion of all mispredicted pixels (D1-all), which were improved by 34 and 33% compared to the second place. Better performance in terms of detail restoration and texture structure clarity was also observed. Detail recovery and texture structure clarity were also enhaced.