Visual IoT Sensing Based on Robust Multilabel Discrete Signatures With Self-Topological Regularized Half Quadratic Lifting Functions

Abstract

The Visual Internet of Things (VIoT) is a next-generation key component to smart cities as it provides more clues for backend analysis. However, the VIoT needs to preliminarily screen/tag data before dispatching them to edge nodes. Moreover, when incomplete data are present, e.g., continuously occluded images, mislabeling may occur, not to mention detecting multilabels accurately. Another challenge is that although pairwise similarities via graph regularization/embedding strengthen label estimation, it is hard to determine the weights/locality between pairwise samples. Besides, static weighting may create rigid topologies owing to fixed distances, which hardly reflect data locality and subsequently cause fitting biases. In view of such, this work presents a discrete signature generation method that can analyze and convert corrupted data into robust signatures, thereby maintaining VIoT recognition functionalities. In this study, half quadratic lifting functions are introduced to accommodate fitting biases caused by data corruption which cannot be effectively solved by $\ell _{2}$, $\ell _{1}$, and $\ell _{2,1}$ norms. To avoid rigid distances among pairwise samples and to enable flexible automatic weighting, self-topological graph embedding/regularization is proposed, where data locality can be dynamically adjusted/learned from data. Additionally, semirelaxation on discrete signatures is devised to avoid quadratic unconstrained binary optimization caused by self-topological graphs and to enable efficient discrete cyclic coordinate descent (DCCD) algorithms for discrete signature optimization. Finally, decorrelation and balancing constraints corresponding to the proposed method are developed. Experiments on open datasets showed that the proposed method yielded better mean average precision (mAP) than the baselines. The outcomes revealed improvements in mAP for distinct noise types, including continuous occlusion, missing values, and sample-specific outliers, with respective increases of $\mathbf {10.97}\boldsymbol{\%}$, $\mathbf {8.05}\boldsymbol{\%}$, and $\mathbf {9.72}\boldsymbol{\%}$, individually. Such findings verified the effectiveness of the proposed idea.