Aiming at the problem of insufficient local and global feature extraction in the feature extraction process of point cloud classification， a point cloud classification algorithm with local-global interaction and structural Transformer was proposed. Firstly， a dual-branch parallel local-global interaction framework was proposed and used to extract local and global features respectively， where in one branch， maximum pooling and convolution were used to extract local features， and in the other branch， global features were extracted by using average pooling and Transformer. Meanwhile， considering the importance of position information in Transformer， a structural Transformer was proposed to further enhance the global structural features by applying interaction of position information with current features for several times. Finally， the local-global features were used for classification to complete the classification task of point cloud. Experimental results show that the classification Overall Accuracies （OAs） of the proposed algorithm are 93.6% and 87.5% respectively on ModelNet40 and ScanObjectNN benchmark datasets. It can be seen that the proposed local-global interaction and structural Transformer network achieve good performance in point cloud classification task.

Concerning the lack of an integrated theory system to select the parameters of combined kernel function used in Support Vector Machine (SVM), one method based on ant colony algorithm and circulated cross validation was put forward to get the optimal parameters. The index named as the mean weighting error was used to evaluate the effect of SVM prediction in different parameters. The value of mean weighting error could be calculated by circulated cross validation. To decrease the calculation workload, the ant colony algorithm was used to enhance the optimization effect of combined kernel function for SVM. This method offered in this paper was applied in the prediction of some plan development cost and the result showed that the optimized combined form of the parameters had the least prediction error. The instance indicates that the parameters optimization method in this paper can improve the prediction precision.

In order to reduce the computational burden and maintain the recognition rate of the image local descriptor, a binary projection method for image local descriptor was proposed. The image patch was projected and transformed into a binary string for boosting the performance as well as speeding up the matching speed. The projection matrix was optimized by machine learning method to maintain its recognition rate and robustness. The experimental result indicates that only a 32-bit binary string is needed to perform as well as the state-of-art descriptors and it shows significantly faster matching speed.

According to the characteristics of clustering with obstacle constraints, using the knowledge of graph theory, a multi-step Arithmetic was proposed. Firstly, it clustered the objects without obstacles by minimum spanning tree clustering method. Then it took obstacles to divide the generated clusters. Lastly it merged the clusters whose obstruct distance was little enough. The algorithm need only one parameter, it is of good performance and can find clusters with arbitrary shapes and varying densities. At last its effectiveness was demonstrated through experiment.