【作者】 陈思；

【导师】 金人超；

【作者基本信息】 华中科技大学 ， 计算机技术（专业学位）， 2022， 硕士

【摘要】 体态检测是一项重要的医学检测,相较于CT图像,通过深度相机获取的深度图像由于其安全又便捷的优点,频繁地出现在体态检测研究中。体表关键点定位是体态检测中极为重要的一步,医师们对深度图像进行手工关键点标注会消耗大量时间且容易引入主观误差。同时,医院的计算机设备往往算力较低,参数量过大的网络会导致运行缓慢,难以被广泛应用。因此,亟需一种轻量级的深度图像精确定位方法。基于OpenPose设计了一种轻量级深度图像体表关键点精确定位网络:FasterOpenPose。针对深度图像带有距离信息的特点,提出了基于三维映射的背景去除方法,将二维坐标映射在三维实际空间中,计算人像边缘的三维坐标,并有效提取了深度图像中的人像区域,降低了杂乱背景对关键点定位的干扰。为满足实际应用中对关键点定位速度的要求,将OpenPose特征提取阶段的图像分类网络VGG-19的前10层,替换为Mobile Net V1的前6层,并使用迁移学习的方法,以Image Net数据集作为数据源域,获取图像分类网络的权重参数,不但减少了特征提取阶段的参数量,而且提高了深度图像特征提取的准确率。其次,裁剪了精炼阶段的向量图预测分支,保留热力图预测分支,减少双分支并行的计算消耗,并将热力图分支的迭代结构改为单支结构。模型深度加深后,会出现模型退化问题,导致定位精度下降,因此本文还设计了级联的深度可分离卷积模块,引入特征融合机制将卷积层输入端和输出端的特征图像进行融合,保留深度图像多阶段的体表特征信息,弥补了网络在前向传播过程中的特征损失,使用该模块代替7×7的大卷积模块,提升了模型的表达能力。Faster-OpenPose模型的总参数量仅为1.48 M,单幅深度图像的平均预测时间为1.31 s,参数量和预测时间优化为OpenPose模型的2.87%和21.20%。在COCO2017数据集上,Faster-OpenPose模型的m AP指标为67.4,AP50指标为88.9,AP75指标为74.2,与其他关键点定位网络的预测精度持平,但模型总参数量和单幅图像的预测时间明显优于其他方法。更多还原

【Abstract】 Body posture detection is one of the most important medical detections.By contrast with conventional CT images,depth images captured by depth cameras were generally utilized in body posture detection research owing to their safeties and conveniences.Further,body surface key point localization is an essential part of posture detection.Manual key point localization in depth images would cost much time of the physicians,and localization results would be easily affected by subjective errors of the physicians.Meanwhile,computer equipments in hospitals commonly have limited computing capability,which will result in that networks with amount of parameters can hardly be handled by the computers and can also hardly be extensively adopted in real engineering.Therefore,it can be said that it is required to develop novel lightweight and accurate depth-image-based localization methods.A lightweight and accurate model is designed based on OpenPose: Faster-OpenPose.Regarding the feature that depth images have abundant three-dimensional information,a novel background removing method is proposed based on three-dimensional mapping.Twodimensional coordinates are mapped into three-dimensional real world coordinate;then,three-dimensional coordinates of body profiles can be calculated.Further,body area can be accurately extracted from the depth images.Complex background of the depth images can be removed,which can greatly increase key point localization accuracy in later analysis.In order to reduce calculation time of key point localization in practical applications,the first10 layers of VGG-19 in feature extraction stage of OpenPose are substituted by the first 6layers of Mobile Net V1;simultaneously,transfer learning method is utilized.Image Net dataset is used as data source domain.Hyper-parameters of the image classification network are fitted using Image Net,which not only reduces the number of hyper-parameters of feature extraction stage,but also improves the accuracy of deep image feature extraction.Subsequently,vector graph prediction branch in the refining stage is trimmed,and heatmap prediction branch is remained.That is,parallel computation of both vector graph and heatmap prediction branches is reduced.After that,to further reduce model computation,iterative structure of the heatmap branch is modified into a single-branch structure.As model depth deepens,model degradation phenomenon occurs,resulting in decrease in localization accuracy.To counter this problem,a cascaded depthwise separable convolution module is designed,and a feature fusion mechanism is adopted to fuse the feature images of input and output of the convolution layer.Additional feature fusion mechanism can preserve multi-stage depth image body surface feature information and compensate excessive feature loss during forward error propagation process.Model expression ability greatly increases by using the cascaded depthwise separable convolution module to substitute 7×7 large convolution module.In total,Faster-OpenPose model only has 1.48 M parameters.It only costs an averaged prediction time of 1.31 s per depth image.Parameter and prediction time decrease by 97.13%and 78.80% compared with original OpenPose model,respectively.In experimental tests of Faster-OpenPose on COCO2017 dataset,its m AP = 67.4,AP50 = 88.9,and AP75 = 74.2.By contrast with other popular key point localization networks,prediction accuracy of Faster-OpenPose is considerable and acceptable;importantly,total number of parameters and prediction time are significantly smaller than those popular key point localization networks.Faster-OpenPose has more application potential in real engineering.更多还原