Research On 2D Human Pose Estimation In Low-resolution Images

2D human pose estimation is a hot topic in computer vision field and has made a breakthrough with the help of deep learning in recent years.However,for low-resolution(LR)images prevalent in real scenes,deep learning based pose estimation methods are far from an acceptable accuracy.The difficulty of pose estimation in LR images is that the information extracted from images is limited and lacks of discrimination.Researches have shown that super-resolution(SR)reconstruction can effectively help other visual tasks to deal with the LR problem.Inspired by this,this dissertation takes SR assisted pose estimation as the basic algorithm framework and reconstructs high-resolution(HR)images that contribute to pose estimation from two aspects,thus improving the accuracy of pose estimation in LR images.In addition,this dissertation combines the pose estimation method in LR images with the pose-based action recognition method to propose a pose-based action recognition method in LR videos.The research results of this dissertation are as follows:(1)General SR methods do not consider regional semantics of the image,resulting in the same level of image quality for the reconstructed foreground and background,which can interfere with the pose estimation network.In order to coarsely enhance the information expression in the body region,this dissertation proposes a body-aware super-resolution(BASR)reconstruction method to assist the 2D pose estimation in LR images.This method combines regional semantics to design a targeted SR loss function.This loss function can guide the SR network to generate HR images with better image quality in the body region than background,thereby weakening the effect of the background on the pose estimation network.The experimental results show that BASR assisted pose estimation obviously outperforms other methods in LR images.(2)In order to further reconstruct fine-grained image features conducive to pose estimation on the basis of BASR,this dissertation proposes a pose-driven super-resolution(PDSR)reconstruction method.This method uses the pose estimation network as a discriminator to drive the SR network to implicitly learn image features that are highly discriminant for pose estimation,and exploits BASR to maintain similarity in pixel space and enhance the information expression in the body region.The experimental results show the effectiveness of PDSR assisted pose estimation.(3)For LR problem in action recognition,this dissertation combines PDSR assisted pose estimation with pose-based action recognition to propose a pose-based action recognition method in LR videos.Based on the two-stream convolutional networks,this method not only extracts robust features by fusing pose information,but also provides still reliable pose sequences in LR videos for networks by using PDSR assisted pose estimation.The experimental results show that this method can obtain a high recognition accuracy for LR video input.