A Diverse Knowledge Perception and Fusion network for detecting targets and key parts in UAV images

Abstract

Detecting targets and their key parts in UAV images is crucial for both military and civilian applications, including optimizing damage assessment, evaluating infrastructure, and facilitating disaster response efforts. Traditional top-down approaches impose excessive constraints that struggle to address challenges such as variable definitions and quantities of key parts, potential target occlusion, and model redundancy. Conversely, end-to-end approaches often overlook the relationships between targets and key parts, resulting in low detection accuracy. Inspired by the remarkable human reasoning process, we propose the Diverse Knowledge Perception and Fusion (DKPF) network, which skillfully balances the trade-offs between stringent constraints and unconstrained methods while ensuring both detection precision and real-time performance. Specifically, our model integrates reasoning guided by three distinct forms of knowledge: contextual knowledge at the image level in an unsupervised manner; explicit semantic knowledge regarding the interactions between targets and key parts at the instance level; and implicit comprehensive knowledge about the relationships among different types of targets or key parts, such as shape similarity. These specific knowledge forms are extracted through a novel adaptive fusion strategy for multi-scale features, a binary region-to-region semantic knowledge graph, and a data-driven self-attention architecture, respectively. Experiments conducted on both simulated and real-world datasets reveal that our method significantly outperforms state-of-the-art techniques, regardless of the number of key parts in the target. Furthermore, extensive ablation studies and visualization analyses validate both the efficacy of our approach and the interpretability of the generated features.