In this study, a method is proposed based on multiscale local and deep features to address the difficulty associated with finding exactly matching pixels from the ill-posed regions in stereo matching. The feature fusion stage comprises two parts. First, the shallow features with different scales, including the Log-Gabor features and the local binary pattern features, are fused. The second part integrates the multiscale shallow fused features and deep features via a convolutional neural network and forms the final feature image, which contains both the semantic and structural information. Further, a positive and negative sample construction method is proposed by adding some noise in the vertical direction to reduce the error that can be attributed to imprecise epipolar alignment in an image. The proposed method is compared with two variant methods (changing or discarding of some modules) with respect to the KITTI datasets. The experimental results validate the effectiveness of the module settings with respect to the proposed method. This method also achieves competitive matching results when compared with those achieved using some classical methods.