融合皮肤检测的卷积姿势机手势分割方法

Convolution pose machine gesture segmentation method fusing skin detection

为解决复杂背景的手势分割问题,提出一种基于融合皮肤检测的卷积姿势机手势分割方法。通过两个CNN网络得到训练的手势分割部分和皮肤分割部分,通过最后一阶段的CNN网络输出最终的手势分割图像,皮肤分割的准确性对最终分割图像起辅助作用,其中核心部分即手势分割部分采用卷积姿势机网络,并运用中间监督的思想将皮肤信息融合。该网络将手势轮廓和经皮肤提取的手势细节结合,分别对轮廓、皮肤、融合3个子网络进行训练,结果对比提取手势的其它方法,验证了该方法的有效性。

To solve the problem of gesture segmentation in complex background,a convolution posture machine gesture segmentation method based on fusion skin detection was proposed.The trained gesture segmentation part and the skin segmentation part were obtained through the two CNN networks,and then the final gesture segmentation image was outputted through the last stage CNN network.The accuracy of the skin segmentation plays an auxiliary role in segmenting the final image.In the gesture segmentation part,a convolutional gesture machine network was used and the idea of intermediate supervision was used to fuse skin information.The network combined the contour of the gesture with the details of the gesture extracted by the skin,and the three sub-networks of contour,skin and fusion were trained respectively.The results were compared with the other methods of extracting gestures,which verified the effectiveness of the proposed method.