A Multi-label Dynamic Learning Framework for Data Trustworthiness Assessment in Mobile Visual Crowdsensing

Mobile visual crowdsensing (MVC) employs images and videos as core data modalities, exhibiting substantial utility in applications such as urban monitoring, scene recognition, disaster response, and environmental surveillance. To address the inherent challenges of image quality and diversity in these domains, we propose a novel MVC framework integrating a multi-label dynamic learning method for assessing data trustworthiness. Our approach evaluates visual crowdsensing data through a tripartite assessment: (1) image semantic similarity, (2) image clarity, and (3) image repetition, while also considering the volume and integrity of user contributions. Experimental evaluations on real-world image datasets within an application system demonstrate that our method effectively assesses the visual quality of sensing data and quantifies user contributions. The framework not only incentivizes users to contribute more images but also ensures high data quality. This work advances MVC research by introducing a dynamic multilabel learning approach for trustworthiness evaluation, with broad applicability to machine learning-compatible visual crowdsensing tasks across various domains.