【摘要】 针对人脸分割的精度问题,提出了融合网络深层特征和浅层特征的新结构,三次精调人脸检测框,提高人脸分割的精确度。新结构结合通道注意力与空间注意力机制,利用深度分离卷积,为每个通道特征提供各自对应的注意力权重,充分利用深层语义信息与浅层定位信息,为精确分割提供特征信息,三次精调为分割提供准确的检测结果。实验结果相比Mask R-CNN的mAP提高0.1,相比最新方法 mAP提高0.2。更多还原

【Abstract】 As regards the precision of face segmentation,a new structure combining the deep and shallow features of the network was proposed,and the face detection frame was fine-tuned three times to improve the accuracy of face segmentation.The new structure combined the mechanisms of channel attention and spatial attention,and utilized depthwise separable convolution to provide corresponding attention weight for each channel feature.And semantic and location information were fully used to provide feature information for precise segmentation,and the third fine-tuning provides accurate detection results for segmentation.Compared with Mask R-CNN,the experimental results of this paper increase mAP by 0.1 and 0.2 compared with the latest method.更多还原