Research On Sign Language Recognition Based On Improved YOLOv3 Algorithm

For people with hearing impairment,sign language is an indispensable way of communication in their daily life.Traditional sign language recognition models include gesture image preprocessing,sign language gesture region detection,sign language gesture feature extraction and sign language recognition.The recognition process is cumbersome and the algorithm complexity is high.Therefore,in the complex background with illumination interference,the recognition rate of traditional sign language recognition model is low,and it is difficult to achieve ideal recognition effect.Aiming at the problems existing in the traditional recognition model,a sign language gesture recognition model based on the improved YOLOv3(You Look Only Once v3)algorithm is established from the application of deep learning method in the field of target detection.Firstly,this paper chooses ten kinds of basic sign language gestures which are common in mathematics: add,minus,multiply,divide,right angle,acute angle,triangle,times,vertical and protractor,and collects these ten kinds of sign language gestures in four different backgrounds.Secondly,the captured video data is converted into a corresponding frame-to-frame image,and 3600 gesture images of each type are selected in the first three backgrounds,i.e.1200 image data in one background,which is used as training set.Then,under the fourth background,500 gesture images of each type,totaling 5000,are selected as the test set.Finally,according to the problems in the preliminary experiments and the shape characteristics of sign language gestures,the detection scale layer of YOLOv3 algorithm has been adjusted four times,which are: the large-scale detection layer is adjusted to small-scale detection layer,the large-scale detection layer is adjusted to medium-scale detection layer,the large-scale and small-scale detection layer is adjusted to medium-scale detection layer,and the large and medium-scale detection layer is adjusted to small-scale detection layer.In the detection layer,Selu function is selected as the activation function of the improved model.For the improved recognition model,firstly,the self-made sign language training set is used to train and test it.The average recognition accuracy of the recognitionmodel is 99.73%.Secondly,the established sign language recognition model is trained and tested using the open Sebastien Marcel gesture data set.The average recognition accuracy is 94.58%.Finally,through the analysis of two experimental data,we can see that compared with the traditional sign language recognition model,the sign language gesture recognition model adopted in this paper has greatly improved the recognition rate,and it has good research value.