Real-time eye tracking technology is the key technology of intelligent eye movement operating system. Compared to the technology based on eye tracker, the technology based on webcam has the advantages of low cost and high universality. Aiming at the low accuracy problem of the existing webcam based algorithms only with the eye image features considered, an optimization technology for eye tracking algorithm with head pose analysis introduced was proposed. Firstly, the head pose features were constructed based on the results of facial feature point tracking to provide head pose context for the calibration data. Secondly, a new similarity algorithm was studied to calculate the similarity of the head pose context. Finally, during the eye tracking, the head pose similarity was used to filter the calibration data, and the data with higher head pose similarity to the current input frame was selected from the calibration dataset for prediction. A large number of experiments were carried out on the data of populations with different characteristics. The comparison experimental results show that compared with WebGazer, the proposed algorithm has the average error reduced by 58 to 63 px. The proposed algorithm can effectively improve the accuracy and stability of the tracking results, and expand the application scenarios of webcam in the field of eye tracking.

Aiming at generalized or fuzzy queries, the content of the returned list of Web search engines is clustered to help users to find the desired information quickly. Generaly, the returned list consists of short texts called snippets carring few information which traditional Term Frequency-Inverse Document Frequency (TF-IDF) feature selection model is not suitable for, so the clustering performance is very low. An effective way to solve this problem is to extend snippets according to a external knowledge base. Inspired by neural network based word presenting method, a new snippet extension approach based on Word2Vec model was proposed. In the model, Top N similar words in Word2Vec model were used to extend snippets and the extended text was able to improve the clustering performance of TF-IDF feature selection. Meanwhile,in order to reduce the impact of noise caused by some common used terms, the term frequency weight in TF-IDF matrix of the extended text was modified. The experiments were conducted on two open datasets OPD239 and SearchSnippets to compare the proposed method with pure snippets, Wordnet based and Wikipedia based feature extensions. The experimental results show that the proposed method outperforms other comparative methods significantly in term of clustering effect.

In order to improve the discriminative ability of dictionary, a dictionary learning algorithm based on Fisher discriminative criterion constraint of the atoms was proposed, which was called Fisher Discriminative Dictionary Learning of Atoms (AFDDL). Firstly, the specific class dictionary learning algorithm was used to assign a class label to each atom, and the scatter matrices of within-class atoms and between-class atoms were calculated. Then, the difference between within-class scatter matrix and between-class scatter matrix was taken as the Fisher discriminative criterion constraint to maximize the differences of between-class atoms. The difference between the same class atoms was minimized when the autocorrelation was reduced, which made the same class atoms reconstruct one type of samples as much as possible and improved the discriminative ability of dictionary. The experiments were carried out on the AR face database, FERET face database, LFW face database and the USPS handwriting database. The experimental results show that, on the four image databases, the proposed algorithm has higher recognition rate and less training time compared with the Label Consistent K-means-based Singular Value Decomposition (LC-KSVD) algorithm, Locality Constrained and Label Embedding Dictionary Learning (LCLE-DL) algorithm, Support Vector Guided Dictionary Learning (SVGDL) algorithm, and Fisher Discriminative Dictionary Learning (FDDL) algorithm. And on the four image databases, the proposed algorithm has higher recognition rate compared with Sparse Representation based Classification (SRC) and Collaborative Representation based Classification (CRC).