CNN联合BI-LSTM混合模型的手势识别算法

Gesture Recognition Algorithm of CNN Combined BI-LSTM Hybrid Model

基于视觉的手势识别是人机交互的热点。针对手势图像易受环境影响造成部分图像缺失,进而导致手势识别精度难以提升的问题,提出了一种基于静态手势图片数据的卷积神经网络(CNN)联合双向长短时记忆循环神经网络(Bi-LSTM)的混合模型。通过随机擦除算法处理手势数据集,可以充分模拟手势缺失问题,首先利用CNN提取手势的空间特征,随后通过Bi-LSTM提取组内图片之间的关联特征,用以解决图片中部分手势图像缺失的问题,最后完成分类。在新加坡国立大学手部姿势数据集上以46.36 f·s^(-1)的速度达到94.6%的识别率,比传统算法识别率更高,鲁棒性更强,有效缓解了因部分手势图像缺失导致识别率不高的问题。

Vision-based gesture recognition is a hot spot in the area of human-computer interaction.To solve the problem that the gesture images are susceptible to environmental influences,which results in the absence of part of the image and makes the gesture recognition accuracy difficult to improve,a hybrid model was proposed,which was based on the static gesture image data of convolutional neural network(CNN)combined with Bi-LSTM.The gesture data set was processed by the random erasure algorithm,which could fully simulate the gesture missing problem.CNN was used to extract the spatial features of the gestures,secondly Bi-LSTM was used to extract the associated features inside a set of pictures to solve the lack of some gestures in the pictures,and finally the classification was completed.The recognition rate was up to 94.6%at the speed of 46.36 f·s^(-1) on the National University of Singapore hand Posture Data set.This method has higher recognition rate and stronger robustness than traditional algorithms,which effectively alleviates the problem of low recognition rate caused by missing part of gesture images.