Research On Finger Language Recognition Method Based On Sparse Coding

Finger language is the language of normal communication and communication between deaf and dumb people.It is a language that coexists with natural language.Its purpose is to achieve convenient,fast and effective communication between deaf and dumb people and normal people and deaf and dumb people.With the continuous development of computer vision technology,intelligent finger language recognition system based on human-computer interaction has increasingly become a research hotspot.Sparse coding can effectively reduce the data processing dimension and increase the processing speed through the sparse expression of the original signal.It is a popular method in the field of computer vision under the background of big data.Therefore,based on the sparse coding method,this thesis has conducted in-depth research on finger language recognition from the four aspects of image acquisition,preprocessing,feature extraction and classification.Firstly,24 types of finger language were collected using a high-definition camera,each sample collected 100 samples,a total of 2400 finger language samples;then,these images were preprocessed,which included threshold segmentation and edge extraction in two parts.Comparison,select the threshold segmentation method with strong real-time performance to segment the image,and select the Sobel operator to extract the image with clear and smooth edges to extract the edge of the image;afterwards,in the feature extraction part,this thesis applies HOG,SIFT,LBP and Gabor method.The method extracts and compares the feature information of each sample of each type of finger language.It is found that the HOG feature extraction method has good invariance to the image geometry and the feature extraction is accurate.Therefore,the HOG method is used to extract the feature of the image;the last in the recognition and classification stage,the feature information of the extracted sample data is mapped to the sparse space,and the complete dictionary is obtained through dictionary learning using the LC-KSVD(Label Consist K-SVD)algorithm,and each type of finger language is sparsely represented,and then combine discriminative sparse coding of discriminant errors with reconstruction errors and categorical errors function,thus achieving effective differentiation of finger language category.In order to verify the effectiveness of the method in this thesis,we tested it on the 24 types of finger language data sets collected by ourselves.Under the selection of different dictionary sizes and different sparse factors,the method of this thesis and the sparse representation classification,D-KSVD SVD),the classification regression tree algorithm and the K nearest neighbor classification algorithm were compared for recognition rate and recognition time.Experimental results show that the proposed method is effective in distinguishing different finger language classifications,with a recognition rate of 99.5% and strong real-time performance.