Anti-loss downsampling and dual-granularity context learning for tiny object detection in remote sensing images

Jie Hu, Jiaming Zhang, Xinbei Zha, Bo Peng, Tianrui Li

Abstract: Object detection in remote sensing images is a critical task for geospatial intelligence. However, it remains particularly challenging when dealing with tiny targets that are characterized by weak feature representation and high sensitivity to complex backgrounds. To address the dual challenges of information degradation during feature compression and insufficient contextual modeling capability, this paper proposes an enhanced architecture that integrates anti-loss downsampling and dual-granularity context learning based on the YOLO framework. Firstly, we introduce a new hybrid downsampling module to replace conventional convolutional downsampling. This module effectively reduces the information loss of tiny objects during feature compression. Secondly, we design a dual-granularity context learning mechanism comprising two complementary components. The first component is a multiscale wavelet convolution module that systematically enlarges receptive fields through cascaded wavelet operators, focusing on capturing local granularity. The second component is a global context aggregation module that employs an attention mechanism to extract global spatial context. Through adaptive weighting, this module strategically aggregates the captured global context into local regions, effectively suppressing background interference. Experimental results show that, on three public remote sensing datasets (AI-TOD, VEDAI, and DIOR), our proposed architecture achieves accuracies of 63.5%, 76.1%, and 85.7%(in terms of mAP50), respectively, which exceed the state-of-the-art methods.

External IDs:doi:10.1007/s10489-025-06874-8