Robust Perceptual Data Sensing Based on Majorization-Minimized Low-Rank Semidefinite Relaxation for Visual IoT

For the Visual Internet of Things (VIoT), terminal devices need to preliminarily screen and tag data for enabling efficient indexing and allocation of sensing data to dedicated edge nodes or cloud data centers for further processing. VIoT terminal devices often face a challenge, where collected data are contaminated by noise or contain partial observed information, which can adversely affect the accuracy and reliability of tagged results. In view of such, this study proposes resilient and perceptual data signature generation for robustly tagging captured data based on low-rank semidefinite relaxation. Such signatures can represent data in a compact way while embedding label information inside at the same time. To increase the robustness and effectiveness of data tagging in the proposed method, nonconvex loss called leaky-minimax concave penalty function is studied. This loss function effectively tackles the challenges posed by partially observed data and instance-pairwise label-space mapping, subsequently improving the reliability and accuracy of the tagging process in the terminal devices. To solve the challenges associated with nonconvex loss functions, this study employs the majorization-minimization technique, which helps conquer the nonconvex optimization problem efficiently. Additionally, low-rank semidefinite relaxation is applied to the formation of data signatures to avoid discrete variable optimization problems caused by Laplacian graph embedding, where data locality learning is performed. The relaxation helps simplify the optimization process and improve the applicability of the method. Experimental evaluations conducted on open data sets have demonstrated the superior performance of the proposed method compared with the baselines, particularly under different data corruption conditions. The results showed an improvement of
$\mathbf {19.45}\boldsymbol {\%}$
,
$\mathbf {8.97}\boldsymbol {\%}$
, and
$\mathbf {5.45}\boldsymbol {\%}$
in F1 scores, with respect to impulse noise, continuous occlusion, and missing values. This validated the effectiveness and practicality of the proposed model in handling VIoT data challenges.