Real-time image alignment: subspace learning based on image gradient direction

one minute per day, taking you through the top conference articles of robots Title: Online Robust Image Alignment via Subspace Learning from Gradient Orientations Author: Qingqing Zheng, Yi Wang, and Pheng Ann Heng Source: International Conference on Computer Vision (ICCV 2017) Compile: blogger Individuals are welcome to forward their circle of friends; other institutions or self-medias need to reprint, the background message application authorization


高鲁鲁的 image alignment has always been a hot topic for scholars. Because of the number of pictures, brightness conversion, partial occlusion, partial pixel damage, etc., the difficulty of image alignment is greatly improved.

In order to solve this problem, this paper proposes a real-time image alignment method, whose core idea is to perform subspace learning on image gradient orientations (IGO). Therefore, the algorithm has three key technologies: subspace learning, deformed IGO reconstruction, and image alignment.

The algorithm is inspired by the PCA-IGO algorithm, and their experiments show that the gradient direction provides a more robust low-dimensional subspace compared to the pixel intensity. Therefore, we have abandoned the traditional method of aligning pixel intensities and then aligning them in the IGO space, which means that the alignment problem of the new image can be decomposed into a linear combination of IGO-PCA based (basis, derived from aligned images). And residual optimization. The solution to this optimization problem is to minimize the L1-Norm of the residual by iteration. Of course, with the addition of new aligned images, IGO-PCA bases need to be iteratively updated. The update method used in this paper is SVD (see the paper for the detailed principle, which is actually the principal component of the gradient).

Finally, this paper validated the effectiveness of the proposed algorithm by performing image alignment and face recognition on a very challenging data set. In a word, the algorithm in this paper has a higher degree of immunity to luminance conversion and occlusion, and is better than other algorithms.

这里写图片描述 图1 The algorithm flow in this article is interested in reading the paper. It is unclear in one sentence and two sentences. In short, the alignment problem becomes a problem of solving linear equations in the gradient direction domain.

这里写图片描述 图2 The first line is the headshot of someone, and the second line is the alignable frame that the algorithm finds from a video stream.

这里写图片描述 图3 This algorithm is used to align prostate MR photographs, where (a) is a number of different subject prostate MR photographs that are not aligned, (b) is a fused photo before alignment, and (c) is a photograph aligned here, ( d) is SIFT, (e) is RASL, and (f) is t-GRASTA. It can be found that the alignment of the algorithm in this paper is particularly good, especially in the blue box.