Focusing on the problems of small samples, large face pose changes, occlusion and complex background, under unconstrained condition, a face verification method based on 3D frontalization and similarity learning was proposed. Firstly, the 3D frontalization progress was applied to generate the frontal face of the face image. Secondly, the complex background was removed by cropping the relevant face regions. Finally, a similarity learning method based on intra-personal subspace was applied to measure the similarity of the image pairs. Experiments were conducted on several databases that were built up by preprocessing the Labeled Faces in the Wild (LFW) database. the difference between these databases and original LFW is their images have been preprocessed. In the experiment with Local Ternary Pattern (LTP) descriptor as the feature extraction method and 625 training image pairs, the recognition rate of the proposed algorithm Similarity Learning over subspace (sub-SL) was 15.6% and 8.4% higher than that of Metric Learning over subspace (sub-ML) and Similarity Metric Learning over subspace (sub-SML) respectively. Experimental results show that the proposed algorithm can effectively improve the accuracy of face verification under unconstrained conditions.

The classificationn decision process of Support Vector Machine (SVM) involves the study of original training samples, which easily causes privacy disclosure. To solve this problem, a classification approach with privacy preservation called IC-SVM (Information Concentration Support Vector Machine) was proposed based on information concentration. Firstly, the original training data was concentrated using Fuzzy C-Means (FCM) clustering algorithm according to each sample point and its neighbors. Then clustering centers were reconstructed to get new samples through information concentration. Finally, the new samples were trained to get decision function, by which classification was done. The experimental results on UCI and PIE show that the proposed method achieves good classification accuracy as well as preventing privacy disclosure.

In order to make full use of depth images and skeleton data for action detection, a multi-feature human action recognition method based on depth images and skeleton data was proposed. Multi-features included Depth Motion Map (DMM) feature and Quadruples skeletal feature (Quad). In aspect of depth images, DMM could be captured by projecting the depth image onto the three plane of a Descartes coordinate system. In aspect of skeleton data, Quad was a kind of calibration method for skeleton features and the results were only related to the skeleton posture. Meanwhile, a strategy of multi-model probabilistic voting model was proposed to reduce the influence from noise data on the classification. The proposed method was evaluated on Microsoft Research Action 3D dataset and Depth-included Human Action (DHA) database. The results indicate that the method has high accuracy and good robustness.

In order to solve the problem that a lot of industrial remote monitoring data transmission would leads to the transmission delay on General Packet Radio Service (GPRS) network, a lossless compression method of industrial remote monitoring data based on improved FPC (Floating Point data Compression) was proposed in this paper. First of all, according to the characteristics of industrial remote monitoring data, the structure of predictor in original FPC algorithm was improved, and then was combined with range encoding algorithm to compress the entire monitoring data. The experimental results show that the prediction precision of improved FPC is higher than before, and the compression ratio is enhanced with the same high compression efficiency. The results of comparison experiments between the proposed method and general lossless compression algorithms show that, the average compression ratio is increased more than 12% and the average compression time is decreased more than 38.5%, which leads to the result that the transmission time is decreased more than 23.7%. The method can increase real-time monitoring performance when network transmission rate is very low and transmission data is very large.

In image segmentation based on clustering analysis, spatial constraints were imposed so as to reduce noise but preserve details. Based on Fuzzy C-Means (FCM) method, a multi-dimensional fuzzy clustering image segmentation algorithm based on kernel metric and local information was proposed to compromise noise and details in the image. In the algorithm, two extra images based on local information derived from the original one through a smoothing and a sharpening filter respectively were introduced to construct a multi-dimensional gray level vector to replace the original one-dimensional gray level. And then kernel method was employed to strengthen its robustness. In addition, a penalty term, which represents the diversity between local pixel and its neighbors, was used to modify the objective function so as to improve its anti-noise ability further. Compared with NNcut (Nystrom Normalized cut) and FLICM (Fuzzy Local Information C-Means), its segmentation accuracy achieved almost 99%. The experimental results on natural and medical images and parameter adjusting demonstrate its favorable advantages of flexibility and robustness when dealing with noise and details.

Most of current segmentation algorithms in H.264/AVC compressed domain lose the local Motion Vector (MV) field and have high time complexity because of global motion compensation. A new fast segmentation algorithm, called Convex Hull segmentation in Spatial-Temporal Filter (STF) based on Boundary Search on compressed domain (BS-CHSTF) was proposed. Motion vector field in bit stream was mainly used in this algorithm. Firstly, the STF algorithm was used in preprocessing the MV, and then eight-direction adaptive search algorithm was used to get connected region, which is filled by constructing the convex hull using boundary of the connected region. Afterwards, multiple connected regions were clustered by redefining the distance between connected regions distance. Finally, the motion object segmentation was obtained by optimizing the mask. Compared with Gaussian Mixture Model (GMM) segmentation algorithm and Ant Colony Algorithm (ACA), the experimental results show that the segmentation accuracy is improved about 3% averagely, and even in the case of lack of motion vector field, the segmentation accuracy increased nearly 20%, while the segmentation speed increased an average of nearly 25%. The method focuses on obtaining the moving object quickly with better segmentation accuracy even in the case of Moving Object (MO) uncompleted.