Research On Hand Gesture Recognition Algorithm Based On Lightweight CNN

As an important human-computer interaction method,hand gesture recognition has always attracted much attention from academia and industry.In recent years,the hand gesture recognition technology based on deep learning has developed rapidly,however,the current hand gesture recognition algorithms have some shortcomings.On the one hand,the current hand gesture recognition algorithms are not good at multihand gesture detection,which makes them difficult to be applied in practical scenarios;on the other hand,the algorithm models generally have the problem of large computational overhead,which makes them difficult to be deployed to devices with limited computational resources.In this paper,to address the above problems,a lightweight method for hand gesture recognition based on deep learning is presented,and a top-down lightweight two-dimensional multi-hand gesture estimation method is proposed.The main research contents including:In the hand localization stage,the lightweight SSD network structure is firstly constructed to achieve model lightweighting by incorporating depth-separable convolution.Secondly,a richer feature fusion is achieved by adaptive spatial feature fusion to effectively improve the network’s detection of small-scale targets.In addition,the algorithm is improved using Soft-NMS in the post-processing stage to address the problem of missed detection in the case of overlapping hand targets.For the proposed method,experimental validation on the PMN dataset shows the improved lightweight SSD model has good performance with a 0.4 percentage point improvement in the m AP value with a significant decrease in the model computational overhead.Meanwhile,the detection accuracy of the lightweight SSD model is higher than other lightweight target detection models.In the hand pose detection stage,a lightweight HRNet network structure is first built,the residual module is improved using depthwise separable convolution,and the number of exchange units used for feature exchange is reduced to achieve model lightweighting.Second,a channel attention mechanism is added to the network to enhance the model’s ability to mine feature information.In addition,lightweight hand pose distillation is proposed and used to solve the problem of insufficient learning capability of the lightweight base network.At the same time,the basic model can maintain high prediction accuracy under the condition that the number of parameters and FLOPs are reduced.Finally,comparison experiments and ablation experiments are designed on the PMN dataset to verify the superiority of the lightweight hand pose distillation method based on the attention mechanism.In this paper,the above lightweight SSD hand localization method and the lightweight hand pose distillation method based on the attention mechanism are effectively fused to build a multi-hand pose estimation model based on cascaded parallel networks.The experimental results show that this model can achieve high prediction accuracy,and can achieve a good balance between the accuracy and the calculation cost of the model.