UCRS: Uncertainty-Calibrated Region Sampling for Real-Time Small Object Detection in High-Resolution Images

Abstract

The detection of objects in high-resolution (HR) images, particularly those taken by surveillance systems or unmanned aerial vehicles (UAVs), appears to be a significant computational bottleneck. Additionally, I observe that when traditional methods process the entire input domain uniformly, the inherent sparsity of objects in backgrounds results in a great deal of redundancy. The Uncertainty-Calibrated Region Sampling (UCRS) framework is presented in this paper. The inefficiency is addressed by the UCRS framework, which uses learning to dynamically allocate computational resources and measure prediction confidence (μ, σ2). Prediction uncertainty is modeled by UCRS using a Gamma distribution. The system can allocate resources as an inference problem thanks to UCRS. The allocation is guided by a Confidence-Weighted Likelihood Metric (Econf ) in UCRS. Inspection of the areas with anticipated detection utility is given priority by UCRS. A test was used to assess UCRS. When compared to processing techniques, the test demonstrates that UCRS reduces floating-point operations (FLOPs). Additionally, the test demonstrates that UCRS outperforms deterministic selection heuristics in terms of precision by 5–8%. Notably, UCRS maintains real-time performance at 34.9 frames per second (FPS) with only 52.5 GFLOPs on the VisDrone dataset, achieving 62.51% AP50 and 36.78% AP. The framework maintains 55.01% APt and 70.25% APs performance at 19.4 FPS on the TinyPerson dataset. These results confirm uncertainty quantification as a reliable and scalable method for building computationally sustainable and flexible perception systems, showing an excellent efficiency-to-accuracy trade-off necessary for limited edge deployment.