Abstract—The automatic detection of granivorous birds in images acquired in rice-field environments is an important challenge for crop protection, yet it remains difficult due to the small object size, scale variations, partial occlusions, and the complexity of natural scenes. We propose REBDetX01, an object detection model based on the CenterNet architecture, integrating a dedicated refinement module in the heatmap prediction branch, referred to as FRM-hm. This module combines a channel recalibration mechanism and dilated spatial attention to improve the organization of heatmap activations and the localization of object centers, without modifying the network backbone or exploiting temporal information. The model is evaluated on the FBD-SV-2024 dataset, characterized by predominantly small flying bird instances and complex backgrounds. Experimental results show that REBDetX01 outperforms the CenterNet baseline, improving Average Precision (AP) from 63.36% to 65.52% (+2.16 points), AP50 from 86.07% to 87.90% (+1.83 points), and Average Precision for small objects (APS) from 61.84% to 63.03% (+1.19 points), thereby indicating its effectiveness for small-object detection. Ablation studies further show that these gains arise from the complementarity between the channel recalibration and dilated spatial attention mechanisms integrated within FRM-hm. Overall, the results highlight the benefit of targeted heatmap refinement for detecting challenging objects in complex natural scenes.
 
Keywords—rice-field environments, object detection, small objects, CenterNet model, attention mechanism
 
Copyright © 2026 by the authors. This is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited (CC BY 4.0).