A novel optimal seam method which aims to stitch those images with overlapping area more seamlessly has been propos ed. Considering the traditional gradient domain optimal seam method and fusion algorithm result in bad color difference measurement and taking a long time respectively, the input images would be converted to HSV space and a new energy function is designed to seek optimal stitching path. To smooth the optimal stitching path, a simplified pixel correction and weighted average method are utilized individually. The proposed methods exhibit performance in eliminating the stitching seam compared with the traditional gradient optimal seam and high efficiency with multi-band blending algorithm.
Toward the unsupervised clustering for color logo images corrupted by noise, we propose a novel framework in which the logo images are described by a model called singular values based region covariance matrices (SVRCM), and the mean shift algorithm is performed on Lie groups for clustering covariance matrices. To decrease the influence of noise, we choose the larger singular values, which can better represent the original image and discard the smaller singular values. Therefore, the chosen singular values are grouped and fused by a covariance matrix to form a SVRCM model that can represent the correlation and variance between different singular value features to enhance the discriminating ability of the model. In order to cluster covariance matrices, which do not lie on Euclidean space, the mean shift algorithm is performed on manifolds by iteratively transforming points between the Lie group and Lie algebra. Experimental results on 38 categories of logo images demonstrate the superior performance of the proposed method whose clustering rate can be achieved at 88.55%.
This paper proposes a novel shape descriptor, called the normalized weighted shape context (NWSC), for feature-based object matching. A shape context as a global characterization descriptor can represent the distribution of points in a set with scale and rotation invariance, but the current technology only provides invariance under translation, scale, and rotation transformations. This paper employs the inertia ellipse of a shape so that the proposed NWSC not only maintains good invariance under scale and rotation transformations but also obtains very robust and accurate matching results under affine transformations. Weights are assigned to each bin of the descriptor to measure its distinctiveness in the matching process. Moreover, a refining approach is proposed to eliminate mismatched features for self-calibration based on the NWSC. Practical experiments are carried out to evaluate its performance, and the results demonstrate that it significantly outperforms the standard shape context. The experiments show that the proposed approach enhances the matching accuracy to a great extent.
The structured light vision system consists of a CCD camera and a digital projector. Calibration of such a vision system plays an important means of accurate 3D reconstruction of a scene. However, the projection model for both the camera and projector is very complicated because of distorted and nonlinear factors in it. It is unlikely to accurately model a camera with only a few parameters even considering some lens distortions. In order to simplify the system calibration and 3D reconstruction, this work presents a new calibration method that is based on neural network and brought forward according to the characteristics of neural network and vision measurement. The relation between spatial points and image points is established by training the network without the parameters of the camera and the projector, such as focus, distortions besides the geometry of the system. The training set for the neural network consists of a variety of lighting patterns and their projected images and the corresponding 3D world coordinates. Such a calibration method has two distinct advantages. It possesses the complicated nonlinear relation between two-dimensional information and three-dimensional information with the neural network, which can include various kinds of distortion and other nonlinear factors during the imaging period. Experiments are carried out to demonstrate and evaluate the procedure. From the result of training we can find out that through the neural network, it may avoid non-linear operation and obtaining the three-dimensional coordinates directly.
This paper presents a method of pattern design for a 3D vision sensor, which is based on the principles of color-encoded structured light, to improve the reconstruction efficiency. Since an ordinary structured light system using an LCD projector needs to take several images (usually 8-12 images) for recovering the 3D scene, as a result its speed is limited and applications are restricted in acquisition of static environment. For dynamic cases, the 3D measurement is desired to only capture a single image. To realize this, a new method is to use a color projector which can be controlled by a computer to generate arbitrary desired color patterns. A problem of the color encoded projection is the unique indexing of the light codes in the image. It is essential that each light grid be uniquely identified by incorporating the local neighborhoods in the light pattern so that 3D reconstruction can be performed with only local analysis of the single image. This paper proposes a method in design of such grid patterns. Experiments are provided to demonstrate the proposed method with two, three, and four different colors. The maximum possible square matrices are illustrated.