Tang Wei, Jia Fangxiu, Wang Xiaoming
College of Mechanical Engineering, Nanjing University of Science and Technology, Nanjing, China.
Front Neurorobot. 2023 Jan 23;16:1042429. doi: 10.3389/fnbot.2022.1042429. eCollection 2022.
It is of vital importance to stitch the two images into a panorama in many computer vision applications of motion detection and tracking and virtual reality, panoramic photography, and virtual tours. To preserve more local details and with few artifacts in panoramas, this article presents an improved mesh-based joint optimization image stitching model. Since the uniform vertices are usually used in mesh-based warps, we consider the matched feature points and uniform points as grid vertices to strengthen constraints on deformed vertices. Simultaneously, we define an improved energy function and add a color similarity term to perform the alignment. In addition to good alignment and minimal local distortion, a regularization parameter strategy of combining our method with an as-projective-as-possible (APAP) warp is introduced. Then, controlling the proportion of each part by calculating the distance between the vertex and the nearest matched feature point to the vertex. This ensures a more natural stitching effect in non-overlapping areas. A comprehensive evaluation shows that the proposed method achieves more accurate image stitching, with significantly reduced ghosting effects in the overlapping regions and more natural results in the other areas. The comparative experiments demonstrate that the proposed method outperforms the state-of-the-art image stitching warps and achieves higher precision panorama stitching and less distortion in the overlapping. The proposed algorithm illustrates great application potential in image stitching, which can achieve higher precision panoramic image stitching.
在运动检测与跟踪、虚拟现实、全景摄影和虚拟游览等众多计算机视觉应用中，将两幅图像拼接成全景图至关重要。为了在全景图中保留更多局部细节并减少伪影，本文提出了一种改进的基于网格的联合优化图像拼接模型。由于基于网格的变形通常使用均匀顶点，我们将匹配的特征点和均匀点视为网格顶点，以加强对变形顶点的约束。同时，我们定义了一个改进的能量函数，并添加了一个颜色相似性项来进行对齐。除了良好的对齐和最小的局部失真外，还引入了一种将我们的方法与尽可能投影（APAP）变形相结合的正则化参数策略。然后，通过计算顶点与最接近该顶点的匹配特征点之间的距离来控制各部分的比例。这确保了在非重叠区域有更自然的拼接效果。综合评估表明，所提出的方法实现了更精确的图像拼接，重叠区域的重影效果显著降低，其他区域的结果更自然。对比实验表明，所提出的方法优于现有的图像拼接变形方法，实现了更高精度的全景拼接，重叠部分的失真更小。所提出的算法在图像拼接中显示出巨大的应用潜力，能够实现更高精度的全景图像拼接。
