Static Gesture Recognition In Complex Background Based On Convolutional Neural Network

Along with development of the information technology,more and more computers are applied in practical life,and affect every aspect of our lives.Serving as the bridge between human and computer,human-computer interaction technology has became pop topic for a long time,especially the natural human-computer interaction such as speech recognition,gesture recognition and face recognition.In recent years,the development of hardware and deep learning technologies has greatly improved the information processing ability of computers,arises a new research upsurge of intelligent human-machine interaction.As an important way of natural human-computer interaction,gesture recognition has many advantages,and is irreplaceable for many applications such as noisy environment and communication between deaf and mute people.CNN(Convolutional Neural Networks)is a key technology of intelligent image processing.Vision-based static hand gesture recognition is a classical research content in gesture recognition and has many mature applications based on CNN technology.Different from the other image recognition,in the static hand gesture recognition,the categories represents some known and relatively fixed hand shapes and the relationship between gestures and background is weak.Through systematically study of static gesture recognition and CNN techniques,the thesis chose FCN(Fully Convolutional Networks)with image segmentation ability as basic model and a static gesture recognition method is proposed.The method combines the known hand shapes of gestures in the classification process,achieved high recognition rate and strong robustness of static gesture recognition in complex background.The main work of the thesis is as follows:Firstly,the thesis studied several classical methods of gesture recognition selectivity.Secondly,the thesis studied CNN related technologies systematically,include the following contents.CNN basic structure,back propagation theory and CNN layers.Analysis and summary of the basic idea of CNN programming include class definitions,data organization,class inheritance and function calls,which based on the MatConvNet Deep Learning Platform.Several classical CNN models and object recognition CNN models were also studied in the thesis.Thirdly,the thesis studied the performance of FCN for static gesture recognition,in order to summarize the unique properties of FCN in static gesture image processing,as the theoretical basis of the thesis,include following contents.The basic structure and principle of classical FCN.Practical studies on FCN characteristics to the hand gesture images with complex background.The optimization of FCN structure are also studied.Fourthly,a static gesture recognition method based on FCN is proposed,the main contents are as follows.Proposed the FCN Double Decision Algorithm,which realized comprehensive utilization of classification and hand shape information by mining the inter-sample and intra-sample information in FCN predictions.Several auxiliary algorithms were also been proposed to improve the performance.The experimental results show that the method achieves a win-win situation of accuracy and robustness,where both the experimental data sets and some similar gesture samples in other data sets were tested.