Application Of Feature Extraction And Convolutional Neural Networks In Gesture Recognition

The gesture recognition based on computer vision is an important research direction in human-computer interaction. It is able to exchange and process information intelligently without supporting of the professional hardware, thus making the gesture operation more natural, simple and user-friendly. Because of people's own different habits, different people will show different shapes of a same gesture, which results in more difficulties of the gesture recognition, therefore, extracting of the robust gesture feature is particularly important in gesture recognition. This paper researchs on the representation and description of robust feature for gesture image, the main contents include the following aspects:(1) In the image pre-processing, it needs to achieve segmentation of gesture subject from the background image. Human's complexion has a feature that the bright and the color of human's skin are separated, so that using modeling method based on color image to detect skin color area in the image; using binary and morphological methods to filter noise and scatter in images and removing analogous skin color region from the image by labeling to give foreground gesture image.(2)In the gesture feature extraction, coming up with complex feature that includes shape and texture features to describe the gesture information. Firstly, taking polygon fitting on the binary contour of the gesture, and by using Hu moment invariants' translation, rotation, and scaling, and good stability advances to describe the shape characteristic of the gesture; and secondly, due to the gesture with rich texture information in the gray-level space, using the gray-level co-occurrence matrix (GLCM) to extract texture feature of gestures; finally combining the complex features with Canberra distance to classify and recognize gesture.(3) Because artificial design features are unitary, and difficult to express a gesture feature all-sidedly, therefore, this paper builds a convolutional neural network model based on gesture's gray-level information with depth study, it uses network to do convolutional and down-sampling operation layer by layer on the image in order to obtain space and structure features of images, like the edge, the feature points, and the texture and enhance the image's integral and partial features,uses partial correlation theory to do sampling operation on images that after convolution, with that we can reduce the amount of data processing network, also make images having characteristic of rotation, translation and scale invariance.This paper studies the feature extraction method based on the shape and texture and convolutional neural network based on gray-level image which used in the gesture recognition. The experiments and contrast show that methods used in this paper have the robustness and effectiveness in different types of gesture feature expression, and the research conclusions of this paper can provide references and basis for the further development of gesture recognition technology.