Aiming at the problems of high time complexity, slow convergence speed and error-prone matching of low coverage point cloud registration, a point cloud registration algorithm based on region segmentation was proposed. Firstly, the volume integral invariant was used to calculate the concavity and convexity of points on the point cloud, and then the concavity and convexity feature point sets were extracted. Secondly, the regions of the feature points were partitioned by the segmentation algorithm based on the mixed manifold spectral clustering, and the regions were registered by the Iterative Closest Point (ICP) algorithm based on Singular Value Decomposition (SVD), so that the accurate registration of point clouds could be achieved. The experimental results show that the proposed algorithm can greatly improve the coverage of point clouds by region segmentation, and the optimal rotation matrix of rigid body transformation can be calculated without iteration. The algorithm has the registration accuracy increased by more than 10% and the registration time reduced by more than 20%. Therefore, the proposed algorithm can achieve fast and accurate registration of point clouds with low coverage.

Aiming at the problem that the traditional Graph Transduction (GT) algorithm is computationally intensive and inaccurate, a semi-supervised classification algorithm based on C-means clustering and graph transduction was proposed. Firstly, the Fuzzy C-Means (FCM) clustering algorithm was used to pre-select unlabeled samples and reduce the range of the GT algorithm. Then, the k-nearest neighbor sparse graph was constructed to reduce the false connection of the similarity matrix, thereby reducing the time of composition, and the label information of the primary unlabeled samples was obtained by means of label propagation. Finally, combined with the semi-supervised manifold hypothesis model, the extended marker data set and the remaining unlabeled data set were used to train the classifier, and then the final classification result was obtained. In the Weizmann Horse data set, the accuracy of the proposed algorithm was more than 96%, compared with the traditional method of only using GT to solve the dependence problem on the initial set of labels, the accuracy was increased by at least 10%. The proposed algorithm was applied directly to the terracotta warriors and horses, and the classification accuracy was more than 95%, which was obviously higher than that of the traditional graph transduction algorithm. The experimental results show that the semi-supervised classification algorithm based on C-means clustering and graph transduction has better classification effect in image classification, and it is of great significance for accurate classification of images.

To improve the retrieval precision of 3D model with the complex surface, a new method based on curvedness feature was proposed. First, the sample points were obtained on the 3D model surface. The curvedness of these points was obtained by computing Gauss curvature and Mean curvature. The curvedness values showed properties of 3D model surface. Secondly, the centroid of the model was set as the center. The coordinate system in which two coordinate axes were the curvedness value and the Euclid distance between the random point and the center was constructed. The distribution matrix of curvedness feature was obtained by computing the statistical number of the sample points in the different Euclid distance. This distribution matrix was the feature descriptor of the 3D model. This descriptor had the property of rotation invariance and translation invariance, which could well reflect the geometric characteristics of complex surfaces. Finally, the similarity between different models was given by comparing the curvedness distribution matrix. The experimental results show that the proposed method can effectively improve the accuracy of the 3D model retrieval, especially suitable for those models with complex surfaces.

Aiming at the problem of quality evaluation of color restoration image for digital protection of faded cultural relics, the objective quality evaluation methods were researched. Combined the computational advantage of Peak Signal-to-Noise Ratio (PSNR) and structure characteristic of human visual feature information entropy, a color image quality evaluation method was proposed based on information entropy of visual features. A quality evaluation function with weights and the corresponding evaluation algorithm process were established, and the weights were determined by normalization method. Then the function value for comparing the similarity between the color restoration image and the reference color image was calculated by using the evaluation algorithm process. The smaller the value was, the higher the similarity was, and the better the corresponding color restoration image quality was, which could be used for the objective judgement of color restoration method. The quality evaluation parameters of four different performance restoration methods were compared. The experimental results show that, the evaluation results are consistent with the subjective perception of human eyes, and the proposed method is effective.

Focusing on the issue that the significant localized characteristics of the influence of skull on the facial surface shape are not fully considered in the existing joint statistical craniofacial reconstruction methods based on Principal Component Analysis (PCA) modeling, which leads to the inadequate description ability of the craniofacial morphological correlation models, by employing these methods and describing the morphological relationship between skull and face, a new craniofacial reconstruction method based on a Partial Least Squares Regression (PLSR) model of local craniofacial morphological correlation was proposed. Firstly, the defects of the joint statistical shape model based on PCA with skull and face as a whole and the advantages of the local morphological correlation model based on PLSR were deeply analyzed. Secondly, by introducing PLSR into the modeling of craniofacial morphological correlation, and based on craniofacial 3D surface model, whose physiological consistent correspondence was established, and classified according to forensic anthropology knowledge, the PLSR coordinate calculation model for each vertex of facial surface was constructed, with those closely related vertex set on skull as its independent variables. Thirdly, with the coordinates of the unknown skull surface model as input values of the coordinate calculation model, the coordinate of each vertex of the predicted face model was acquired, from which the predicted face could be reconstructed, and the concrete procedure of the new reconstruction method was elaborated. Finally, several craniofacial reconstruction experimentations by applying the new reconstruction method based on PLSR were given, and the new reconstruction method was comparatively analyzed and evaluated by the indicators including effectiveness of reconstruction and absolute error. The experimental results show that the new reconstruction method significantly improves the accuracy of craniofacial reconstruction.

In order to automatically register the skulls that differ a lot in pose with the reference skull, or miss a large part of bones, an automatic nonrigid 3D skull registration method based on boundary correspondence was proposed. First, all the boundaries of target skull were calculated, and according to the edge length and the shortest distance between the edges, the edge type was identified automatically, and the correspondence between the registered skull and the reference skull was established. Based on that, the initial position and attitude of the skull were adjusted to realize the coarse registration. Finally, Coherent Point Drift (CPD) algorithm was used twice to realize the accurate registration of two skulls from the edge region to all regions. The experimental results show that, compared with the automatic registration method based on Iterative Closest Point (ICP) and Thin Plate Spline (TPS), the proposed method has stronger robustness in pose, position, resolution and defect, and has more availability.

Aiming at the error accumulation problem in the process of the traditional global matching of the three-dimensional (3D) models, a global optimal matching method based on swarm intelligences was proposed. The global matching process for multiple 3D fragments was abstracted, and then a mathematic model of the global optimal matching was set up, the solution of the optimal matching for multiple 3D fragments was converted to satisfy certain constraint conditions of the optimal match matrix of combinatorial optimization problem. A discretization algorithm based on hybrid social cognitive optimization algorithm was proposed to solve the NP (Non-deterministic Polynomial) problem. Finally, the classical example analyses verified that the proposed algorithm has global optimization ability and strong robustness without the initial position, and it provides an efficient method for global matching of the 3D fragments.

Aiming at the problem of object-based image retrieval, a novel algorithm named KP-MIL was proposed, which worked in the Multiple Instance Learning (MIL) framework. Firstly, this algorithm clustered the instances in positive set and negative set, and found the potential positive instance and bag structure. Then an alpha coefficient was introduced to trade off between positive instance and bags similarity. Experiments on SIGVAL dataset show that this algorithm is feasible, and the performance is superior to other MIL algorithms.