Highly discriminative feature expression for non-rigid shape recognition is an important and challenging task, which requires both abstract and robust shape descriptors. However, the majority of existing low-level descriptors are designed via hand-crafted, which are sensitive to local changes and larger deformation. To address this issue, this paper proposes a bag of shape descriptor based on unsupervised deep learning and Bag of Words (BoW) for shape recognition. Different from existing pipelines, our method is specially designed to learn high-level and hierarchical shape features from multi-scale context structures. It effectively overcomes obstacles, such as irregular topology, orientation ambiguity, and rigid or non-rigid transformation in the hierarchical learning of contour fragments. Specifically, by adopting an improved decomposing strategy, the shape can be decomposed to a series of valuable contour fragments, results in local to global feature learning. An unsupervised learning framework is also applied to the contour fragment for its feature expression based on the context structure and SSAE (Stack Sparse Auto Encode). In the process of shape representation, a high-level shape dictionary is learned by K-clustering to achieve discriminative feature coding. In addition, to achieve a compact and simplified shape representation, SPM (Spatial Pyramid Matching) is adopted by max-pooling, which effectively incorporates spatial layout information of the given shape. The experiments demonstrate that the proposed method achieves state-of-the-art performance on several public shape datasets comparing with the latest approaches. Our method also obtains high performance under the noisy and occlusion condition.

用于非刚性形状识别的高度区分性特征表达是一项重要且具有挑战性的任务，它需要抽象和鲁棒的形状描述符。但是，大多数现有的低级描述符是通过手工设计的，它们对局部变化和较大变形敏感。为了解决这个问题，本文提出了一种基于无监督深度学习的袋形描述符和用于形状识别的词袋（BoW）。与现有管道不同，我们的方法经过专门设计，可以从多尺度上下文结构中学习高级和分层形状特征。它有效地克服了轮廓片段的分层学习中的障碍，例如不规则拓扑，方向模糊性以及刚性或非刚性变换。具体来说，通过采用改进的分解策略，可以将形状分解为一系列有价值的轮廓片段，从而实现局部到全局特征学习。基于上下文结构和SSAE（Stack Sparse Auto Encode，堆栈稀疏自动编码）的特征表达，无监督学习框架也被应用于轮廓片段。在形状表示的过程中，通过K聚类学习高级形状字典以实现判别性特征编码。另外，为了实现紧凑和简化的形状表示，最大池采用SPM（空间金字塔匹配），它有效地合并了给定形状的空间布局信息。实验表明，与最新方法相比，该方法在多个公共形状数据集上均具有最先进的性能。