文章摘要

文章信息/Info

Title:: Handgesture recognition method based on network-embedded fusion convolution neural network

Author(s):: FENG Xiang; WU Han; SI Bingling; JI Chao; Weifang Medical College, College of Biological Science and Technology, Weifang 261000,China

Keywords:: Convolutional neural network; Network-embedded structure; Information fusion; Feature extraction; Pyramid sampling

摘要:: 手势识别是人机智能交互领域的研究热点。本研究基于LeNet-5网络和信息融合思想提出新的嵌网融合-卷积网络结构来处理手势图像识别问题，新网络在传统卷积网络的卷积层中以多层感知器替换传统线性卷积核来构造特征提取框架，并在卷积层的输出端级联Inception网络结构，同时将金字塔采样机制引入池化层以替换常规随机采样和最大值采样，利用金字塔多尺度融合策略来拼接不同维度的特征，进而将融合特征传输给全连接层，最后在全连接层中引入支持向量机思路进行特征识别。实验仿真中，本研究识别网络在MNIST数字集及自建手势数据集进行验证，识别准确率最高达到98.2%，优于几种常规网络。

Abstract:: Gesture recognition has always been the hotspot in intelligent human & computer interacting field. We proposed a novel network-embedded convolution network structure based on LeNet-5 network and information fusion, to solve the problem of gesture-image recognition. Firstly, the multi-layer perceptron was used to replace the traditional linear convolution kernel to extract the special features. Then, the inception model was cascaded behind the convolution layer. At the same time, the pyramid sampling mechanism was introduced into the pooling layer to replace the conventional random sampling and maximum sampling. The pyramid multi-scale fusion strategy was used to splice features of different dimensions, and then transmitted to the fully connected layer. Finally, the support vector machine idea was introduced into the fully connected layer for recognition. In simulations, the proposed network has been tested in MNIST set and self-built gesture sample set, and the final recognition accuracy is up to 98.2%, which is better than that of several existing conventional networks.