基于改进AlexNet的双模态握笔手势识别

Bimodal pen-holding gesture recognition based on improved AlexNet

本文提出了一种基于改进AlexNet的双模态握笔手势识别方法。该方法根据握笔手势特征自建了 8 100张握笔手势数据集,对数据集进行了手势分割获取二值图像、骨架提取获取包含原图的骨架图像等处理,并将处理后的2种类型图像构成双模态图像输入至改进的AlexNet中。针对AlexNet提取握笔手势特征不充分的问题,本文将AlexNet第一层的卷积核大小修改为3×3,并在卷积层之后添加了批量归一化、注意力机制。通过实验证明,该方法对9种握笔手势的平均识别率达到75.6%,分别高于骨架图像、分割图像、AlexNet网络11%、16%和13%,证明了该模型对握笔手势识别的有效性。

In this paper,a new method of pen-holding gesture recognition based on improved AlexNet is proposed.In this method,8100 pen-holding gesture data set is constructed according to the characteristics of pen-holding gesture.Then the data set is processed,including gesture segmentation obtains binary image,skeleton extraction obtains the skeleton image containing the original image,and the processed images are formed into bimodal images which are input into the improved AlexNet.In order to solve the problem that AlexNet is not sufficient in extracting the characteristics of pen-holding gesture,this paper modifies the convolution kernel size of AlexNet's first layer to 3×3,meanwhile adds batch normalization and attention mechanism after the convolution layer.The experimental results show that the average recognition rate of nine pen-holding gesture is 75.6%,which is 11%,16%and 13%higher than that of skeleton image,segmenting image and AlexNet network,respectively.It proves the effectiveness of the proposed model for pen-holding gesture recognition.