多尺度卷积特征融合的 SSD 手势识别算法-《计算机技术与发展》

文章信息/Info

Title:: SSD Gesture Recognition Algorithm with Multi-scale Convolution Feature Fusion

Author(s):: XIE Lin-dong; ZHONG Zhi-dan; QIAO Dong-hao; GAO Xin-hong; School of Mechanical and Electrical Engineering,Henan University of Science & Technology,Luoyang 471003,China

Keywords:: multi - scale convolution features; small - medium proportion gesture; dilated convolution; deconvolution; feature fusion; improved loss function

摘要:: 为了提高对中小占比手势识别的准确性与稳定性,提出了一种多尺度卷积特征融合的 SSD(single shot multibox detector)手势识别方法。该方法突出表现在两大方面,其一,在原始的 SSD 算法的多尺度卷积检测方法基础上,引入了不同卷积层的特征融合思想,经过空洞卷积下采样操作与反卷积上采样操作,实现网络结构中的浅层视觉卷积层与深层语义卷积层的融合,代替原有的卷积层用于手势识别,以提高模型对中小目标手势的识别精度;其二,为了解决正负样本不均衡导致分类性能差的问题,提出一种改进的损失函数,以提升模型对目标手势的分类能力。在手势识别公开的数据集上的实验结果表明,与 SSD 和 Faster R-CNN 等识别方法相比,能够在保持较高的手势检测精度的同时,又具有较好的鲁棒性与检测速度。

Abstract:: To improve the accuracy and stability of small-medium proportion gesture recognition,SSD (single shot multi-box detector) gesture recognition algorithm with multi-scale convolution feature fusion is proposed. Two asp-ects are highlighted in this method. On the one hand,based on the multi-scale convolution detection method? of the original SSD algorithm,the feature fusion mechanism of different classification layers is introduced. Through the dilated convolution down sampling operation and the deconvolution up sampling operation,the shallow visual feature layer and the deep semantic feature layer in the network structure are organically combined to replace the original convolution layer for gesture recognition to improve the semantic represent-ation ability of the model. On the other hand,to solve the problem of poor classification performance caused? by imbalance of positive and negative samples,an improved loss function is proposed. Experiments on the open data set of gesture recognition show that compared with SSD,Faster R-CNN and other recognition methods,the proposed method has better robustness and detection speed while maintaining higher gesture detection accuracy.