Research On Chinese Sign Language Recognition Based On Object Detection

The number of deaf-mute people in China accounts for a large proportion of the total population of the country.Sign language is an important tool for communication between deaf-mute people and normal people.The significance of studying sign language recognition is not only to improve communication between deaf-mute people.It can help normal people to understand the sign language expressions of deaf and mute people,and can help normal people to better help deaf and mute people,and deaf and mute people can better communicate with society.In this paper,some methods of deep language recognition have been studied with the help of some methods in deep learning.In traditional image recognition,the model used is generally relatively complex in structure.Although the pooling layer can extract information-related weight coefficients well,the structure is heavy It is difficult to calculate.The purpose of this article is to build a lightweight model that enables non-deaf people to perform real-time sign language recognition on mobile devices such as mobile phones.The research process of this paper is as follows: First,the static sign language is recognized,and the focus is on the static sign language.In the experiment,YOLO-LITE in the YOLO series was used to recognize the sign language.Secondly,analyze the problems in the identification of the model,and improve the model so that the improved model can basically meet the conditions of real-time detection to achieve the purpose of the experiment.The network model used in this paper has a simple network structure and few layers,especially it can be used on GPU-free computers or mobile devices,which meets the needs of mobile devices in this experiment to a certain extent.In the preparation of this experiment,I found that there are not many data sets about sign language.So I decided to collect the data set of this experiment by myself.To ensure the authenticity and originality of the experimental data.Transfer learning was carried out on the parameters during the experiment,and iterative training on transfer learning was performed to ensure the rationality of the parameters,to a certain extent,the accuracy of the experiment can be improved,and the data set of the University of Science and Technology of China In comparison,the structure of Tiny-YOLOv2 and YOLO-LITE,which are also lightweight models in YOLO,are very similar in structure,so Tiny-YOLOv2 was selected for a comparative experiment.The experimental results show that YOLO-LITE can achieve the speed of sign recognition Quick identification.After many experiments,the results show that: The experimental speed of YOLO-LITE has been improved to a certain extent,but the accuracy of the experiment is not high,and it cannot meet the needs of real life.The recognition rate is very high,and the accuracy of feature extraction is very high,but the overall structure has too many layers.The recognition rate of the anti-residual convolution module of the lightweight Mobile Netv2 is very high.In the improvement of the lightweight Mobile Netv2 anti-residual convolution module,the results show that the accuracy rate has been greatly improved,while also ensuring the advantage in speed,the accuracy rate has also been improved accordingly,can be achieved 92%,an increase of 12%compared to before the improvement.