基于改进的Faster RCNN的手势识别

Gesture Recognition Based On Improved Faster RCNN

针对传统手势识别算法存在手势定位不精确,识别率不高,鲁棒性不强等问题,提出改进的Faster RCNN网络进行手势的精准定位和识别。Faster RCNN采用强语义信息、低分辨率的顶层特征图作为RPN网络的输入,导致对小目标识别率不高。改进的Faster RCNN结合FPN网络算法,将高层特征通过上采样不断与前层特征融合,构造不同尺度的特征金字塔模型作为RPN网络的输入,提升了Faster RCNN对手势的检测效果。

Aiming at the problems of traditional gesture recognition algorithms such as inaccurate gesture positioning, low recognition rate and low robustness, an improved Faster RCNN network is proposed to accurately locate and identify gestures. Faster RCNN uses strong semantic information and low-resolution top-level feature maps as input to the RPN network, resulting in low recognition rate for small targets. The improved Faster RCNN combines the FPN network algorithm to integrate the highlevel features with the previous features through up-sampling, and constructs the feature pyramid model with different scales as the input of the RPN network, which improves the detection effect of Faster RCNN on gestures.