| Summary: | Amid advancements in aerospace technology and remote communication, a proliferation of Earth-observing satellites has been launched, creating a distinction between high- and low-altitude platforms. High-altitude satellites capture low-resolution (LR) aerial images, covering expansive areas, whereas their low-altitude counterparts provide high-resolution (HR) images of relatively confined spaces. The task of semantically categorizing LR aerial imagery is pivotal within numerous artificial intelligence (AI) systems but is encumbered by challenges including limited availability of labeled training data and the complexity of approximating human environmental perception through computational models. To address these challenges, this research proposes a novel strategy that marries active perception learning with Hessian-regularized feature selection (HRFS). This approach endeavors to procure perceptually and discriminatively potent visual representations for classifying LR aerial imagery. By emulating the human propensity to sequentially engage with salient regions within a visual field, an active learning paradigm is adopted to differentiate between salient and non-salient regions within LR aerial images. Theoretically, this methodology ensures that selected salient regions can reconstruct the aerial imagery in its entirety, thus mirroring the human visual system’s perception. Following this, a pioneering HRFS technique is devised to extract premium features from these selectively identified salient regions, distinguished by its semi-supervised operation, the capability for concurrent linear classifier training, and the preservation of the geometric distribution of samples within the feature space. Empirical assessments underscore the resilience and efficacy of the proposed classification framework.
|
|---|