【作者】 胡朝阳；

【导师】 汪国有；

【作者基本信息】 华中科技大学 ， 模式识别与智能系统， 2019， 硕士

【摘要】 随着深度学习方法的不断发展与成熟,基于深度卷积网络的语义分割已经成为自动驾驶、室内导航、遥感制图等视觉感知领域应用的一项重要技术。然而对于弱小或模糊的目标,语义分割方法仍然存在分类精度低的问题。为了提高语义分割的精度,我们开展了如下两个方面的研究工作:受可变形卷积的启发,舍弃普通卷积的固定采样,提出了可变向卷积网络,达到更加灵活的可变向的采样。通过将特征图上的每一个点的邻域分为五个不同方向的区域,并分别提取五个区域的多尺度局部特征,同时预测五个区域特征的权重因子,然后通过softmax软约束对权重因子重新赋值,放大可分特征,抑制干扰特征,最终达到语义特征更显著的目的。虽然可变向卷积增加了少量计算量,但是同时为模型融入了更加可分的多尺度局部上下文特征,相比于原始的FCN(Fully Convolutional Networks),我们的算法取得了更高的精度。提出了微U型网络的残差编码及动态路由编码的全局上下文分割算法。在原有的U型网络基础上,采用更少采样层数的简化U型网络,提取的是粗粒度的语义特征,但是引入残差编码和动态路由编码两种策略分别对场景的全局上下文进行无监督聚类,获取更具表征性的语义并替代原有通过全局池化获得的语义。虽然一定程度上增加了计算量,但是相比于FCN,我们的算法精度更高。我们分别在Pascal VOC2012和Pascal Context数据集进行了大量的实验,实验证明,我们的算法极大地提高了对象的分割效果。与现有的经典算法对比,我们的算法取得了更优的结果。更多还原

【Abstract】 With the continuous development and maturity of deep learning methods,semantic segmentation based on deep convolutional networks has become an important technology for visual perception applications in the fields of autopilot,indoor navigation and remote sensing mapping.However,for weak or fuzzy targets,the semantic segmentation method still has the problem of low classification accuracy.In order to improve the accuracy of semantic segmentation,we have carried out the following two aspects of research workFirstly,inspired by the deformable convolution,the fixed sampling mode of ordinary convolution is discarded,and a variable direction convolution network is proposed to achieve more flexible variable-direction sampling.By dividing the neighborhood of each point on the feature map into five different directions,and extracting the multi-scale local features of the five regions,and simultaneously predicting the weighting factors of the five regional features,and then weighting factor is re-assigned by using the softmax soft constraint.So the weight of the salient features is magnified and the interference features are suppressed so that the semantic features become more salient.Although the variable convolution adds a small amount of computation,multi-scale local context features are incorporated for the model,our algorithm achieves higher precision than the original FCN(Fully Convolutional Networks)Secondly,a global context segmentation algorithm for residual coding and dynamic routing coding based on micro U-shaped networks is proposed.On the basis of the original U-type network,a simplified U-shaped network with fewer sampling layers is adopted,and the coarse-grained semantic features are extracted.However,two strategies,residual coding and dynamic routing coding are introduced to unsupervised clustering of the global context of the scene,and more representative semantics are obtained instead of the original semantics obtained through global pooling.Although the computational complexity is increased to some extent,our algorithm is more accurate than FCN.Although the amount of calculation is increased to some extent,our algorithm is more accurate than FCNWe have done a lot of experiments on Pascal VOC2012 and Pascal Context datasets respectively.The experiments show that our algorithm greatly improves the segmentation effect of objects.Compared with the existing classical algorithms,our algorithm achieves better results.更多还原