Research On Sign Language Recognition Algorithm Based On Improved YOLOv5s

Sign language is the main communication tool used among hearing-impaired individuals and between hearing-impaired individuals and hearing individuals.In realistic sign language recognition tasks,factors such as complex background environments can increase the difficulty of detection,leading to unstable convergence of traditional network models and low detection efficiency.To solve the communication difficulties between hearing-impaired and hearing individuals and to address the shortcomings of traditional recognition networks,this article proposes an improved sign language recognition network based on the YOLOv5 s network model.The main work is as follows:(1)To address the insufficient feature extraction and low accuracy of complex background targets in current sign language recognition tasks,this article improves the channel domain of the CBAM(Convolution Block Attention Module)attention mechanism,solves the problem of channel information loss caused by dimensionality reduction,and proves through experiments that integrating the improved CBAM module into the YOLOv5 s backbone network can most effectively enhance the feature information of detection targets,allowing for more accurate localization and identification of key targets.(2)To address the problems of large deviation in sign language gesture positioning and unstable convergence of traditional network models,this article first applies the K-means++algorithm to improve the size matching of prior anchor boxes,determines the optimal prior anchor box size for the data set used in this article,and achieves accurate matching between prior anchor boxes and actual objects.Secondly,the Lovasz-Softmax Loss and YOLOv5 s original loss function Cross Entropy Loss are weighted and combined to make the network converge more stably during model training and achieve a certain improvement in accuracy.(3)Sign language recognition experiments based on the improved YOLOv5 s were conducted.The improvement methods of attention mechanism,anchor box calculation method,and loss function were combined,and the experimental results show that the average precision value(mean Average Precision,m AP)of the improved network model reached97.67%,which is 3.17 percentage points higher than the original YOLOv5 s model.Moreover,the precision(P)and recall(R)were also improved by 3.44 and 1.89 percentage points,respectively,effectively improving the detection performance of the sign language recognition network.