The purpose of graph matching is to find the correspondence between nodes of two graphs. Existing graph matching models only consider the similarity between nodes when performing cross-graph convolutions, ignoring the consistency of the structure. To solve this problem, this paper proposes a new graph matching model that incorporates neighbor matching into the cross-graph convolution module. Neighbor matching calculates the attention weights for cross-graph convolution based on the local topology of nodes and neighbor dissimilarity. Because different neighbors have differing importance to the central node, neighbor matching first assigns aggregate weights for different neighboring nodes based on feature correlations. The cross-graph neighbor matching model then captures the distinctions among neighbors. Finally, the attention weights of cross-graph convolutions are jointly determined by the similarity between nodes and the consistency of their neighbors. This paper conducts comparative experiments on two public datasets. The experimental results show that on the Pascal Visual Object Classes(Pascal VOC) dataset, compared with the baseline model, the matching accuracy of the proposed model on 20 categories is increased by 0.9% on average; on the Spair-71k dataset, the average accuracy is increased by 1%.

Graph similarity computation is an important problem for research in the field of complex networks, which can further facilitate tasks such as graph classification, clustering and similarity search. Graph similarity is usually measured by the graph edit distance (GED) metric; however, the exact computation of GED is an NP-hard problem with high computational complexity and difficult to solve. In recent years, graph similarity computation using graph neural networks (GNN) has emerged to achieve efficient metric results. To fully exploit the deep information in the graph and obtain more accurate graph similarity computation results, we propose a multidimensional graph matching network model using graph topological information. Firstly, to capture the rich fine-grained information in the graph, a multidimensional graph matching module is proposed in the model, including cross-graph feature interactions at the node-graph level as well as at the multi-level graph-graph level, and the expressiveness of the model is improved by the graph matching module. Secondly, graph topology feature matching is added in the similarity calculation to focus on how similar a pair of graphs are in terms of topology and to utilize topology information more fully. We conducted experiments on real-world datasets to demonstrate the effectiveness of the model.

Automatic video mosaicking is a challenging task in computer vision. Current researches consider either panoramic or mapping tasks on short videos. In this paper, an automatic mosaicking algorithm is proposed for both mapping and panoramic tasks based on the adapted key-frame on videos of any length.The speeded up robust features（SURF） and the grid motion statistic（GMS） algorithm are used for feature extraction and matching between consecutive frames, which are used to compute the transformation. In order to reduce the influence of the accumulated error during image stitching, an evaluation metric is put forward for the transformation matrix. Besides, a self-growth method is employed to stitch the global image for long videos. The algorithm is evaluated by using aerial-view and panoramic videos respectively on the graphic processing unit（GPU） device, which can satisfy the real-time requirement. The experimental results demonstrate that the proposed algorithm is able to achieve a better performance than the state-of-art.

The Histogram of Oriented Gradient (HOG) feature for pedestrian detection has achieved good results, but it is time-consuming. For resolving this problem, a modified method for HOG is proposed to reduce the dimension of the features. On the base of analyzing the process of HOG, nine independent HOG channels (HOG-C) are extracted according to the gradient orientation interval. Through evaluating the effectiveness of HOG-C for pedestrian detection individually, a combination of HOG channels (CHOG-C) feature is presented based on statistical regularities. Comprehensive experiments on INRIA database demonstrated the promising performance of the CHOG-C feature, and the experimental results shown that the dimension is reduced meanwhile without losing the accuracy.