Research On Real-time Recognition Method Of Static Gestures Based On Deep Learning

As a simple and easy-to-express communication method,gestures have very important research value no matter in the field of human-computer interaction or in other application fields.Compared with the shortcomings of traditional hand-designed features such as cumbersome and complex,low recognition rate,deep learning methods can automatically extract image features,and are less affected by image background factors,which greatly improves the recognition rate.However,with the development of target detection technology,most algorithms have been continuously increasing the depth of the network in order to obtain a higher recognition rate,but have ignored the large number of parameters and the large amount of computing resources that are brought about by doing so,thereby reducing The recognition speed of the model and the portability of the model on mobile devices are improved.In response to the above problems,this research designed a real-time recognition model for static gestures based on deep learning.While ensuring the balance of model detection accuracy and speed,the training time of the model and the model’s memory ratio are reduced as much as possible,making it beneficial in Deployment on mobile devices.Aiming at the characteristics of limited computing resources and small storage space of mobile devices,an efficient Shuffle Netv2 and YOLOv3 integrated network static gesture realtime recognition method is proposed to reduce the demand for hardware computing capabilities of the model.In order to solve the problems of the traditional YOLOv3 model’s backbone network Darknet-53 with large parameters and large memory usage,a lightweight network Shuffle Netv2 was introduced to replace the original backbone network to reduce the computational complexity of the model,and then in the model The CBAM attention mechanism is added to strengthen the network’s attention to the channel and space,and finally the K-means algorithm is used to re-cluster the a priori frames of the self-made data set,so that it can accurately locate the target to improve the detection accuracy of the model.The experimental results show that the improved algorithm guarantees the balance of detection accuracy and detection speed,and at the same time,the training time of the model and the weight ratio of the memory have been greatly improved.When the PASCAL VOC dataset is used to test the above-mentioned improved model,it is found that if there are dense targets in the image,the improved model will have the problem of missed detection.In response to this problem,an Anchor-Free-based target detection algorithm Center Net is used,which does not require additional calculation of the prior frame size,which not only reduces the amount of calculation,but also achieves good results in intensive target image detection.Similarly,based on mobile devices,Center Net also occupies a large amount of computing resources.Here,a lightweight network designed for mobile devices,Mobile Netv3,is introduced to replace the backbone network of the Center Net algorithm to optimize the original algorithm.Finally,the experimental results show that the optimized model has achieved good results,which is conducive to deployment on mobile devices.