小波变换低频信息与Xception网络的静态手势识别

Low-frequency Information of Wavelet Transform and Static Gesture Recognition of Xception Network

轮廓对于提高手势识别的准确率与缩短响应时间具有重要作用。经过小波变换得到的低频信息能准确反映手势轮廓。Xception卷积神经网络能减少模型参数并获得更高的分类准确率。因此,提出一种小波变换低频信息与Xception网络的静态手势识别方法。首先,将原始ASL的8000张手势图像经二维小波变换批量处理后得到低频、水平高频、垂直高频和对角线高频共4种图像,然后将低频图像作为Xception的输入进行手势识别,并设计了原始、水平高频、垂直高频和对角线高频4种图像的对比实验。实验结果表明,Xception能对低频信息与原始信息进行有效的特征学习,低频图像内存是原始图像的2/13,运行时间是原始图像的88.5%,但低频图像的准确率只比原始图像低0.2%。该方法大大减少了训练图像所需的存储容量并提高了运行速度。最后比较Xception与VGG16、VGG19、ResNet和ResNetV2常用的手势识别网络,表明Xception在较短时间内能取得更好的识别效果。

Contours play an important role in improving the accuracy and response time of gesture recognition.The low-frequency information obtained by wavelet transform can accurately reflect the contour of the gesture.Xception convolutional neural network can reduce model parameters and obtain higher classification accuracy.Therefore,a static gesture recognition method based on wavelet transform low-frequency information and Xception network is proposed.First,the 8000 gesture images of the original ASL are batch-processed by two-dimensional wavelet transformation to obtain 4 kinds of images:low frequency,horizontal high frequency,vertical high frequency and diagonal high frequency,and then use the low frequency image as the input of Xception for gesture recognition.In order to verify the effectiveness of low-frequency information,a comparative experiment of four images of original,horizontal high frequency,vertical high frequency and diagonal high frequency was designed.Experimental results show that Xception can perform effective feature learning on low-frequency information and original information.The low-frequency image memory is 2/13 of the original image,and the running time is 88.5%of the original image,but the accuracy of the low-frequency image is only 0.2 lower than the original image.This method greatly reduces the storage capacity required for training images and increases the running speed.Finally,a comparison between Xception and VGG16,VGG19,ResNet and ResNetV2 commonly used gesture recognition networks shows that Xception can achieve better recognition results in a shorter period of time.