Abstract: Based on the available algorithm, it is a tough work to achieve stitching of wide-baseline weak-texture images with parallax discontinuity. As a result, the stitching task usually requires manual intervention. For this, we modify the critical steps of image matching and image registration, and propose an anti-parallax automatic stitching algorithm for wide-baseline weak-texture images. First, we obtain the quasi-dense correspondence of weak-texture features from coarse to fine, based on the local feature transformers model incorporating the geometric correction of the image perspective. Next, based on matching points and deep neural network (DNN), the reliable perspective transform between wide-baseline images can be learned to eliminate global registration disparity, and then the local left disparities are precisely fitted by thin plate spline (TPS) function. Furthermore, the polygon boundary of the image stitching result is regularized, and it is trained as a regularized rectangle through a fully convolutional network, which effectively removes the blank area and preserves the content of the image stitching to the maximum extent. Finally, four groups of UAV and ground close-range wide-baseline stereo image pairs with weak-textures are selected and tested, and the results of image matching and registration stages of our method are respectively compared with the results of the existing representative algorithms. The experimental results verify that our method has significant advantages in the number of matching points, accuracy and stitching quality, and show good stability at the weak-texture and parallax discontinuity regions of the images.

Key words: image matching, disparity discontinuity, wide-baseline weak-texture images, deep neural network, automatic stitching