基于特征增强与损失优化的密集鸟群计数算法

doi:10.17520/biods.2026005

生物多样性

基于特征增强与损失优化的密集鸟群计数算法

刘国辉^1,2, 胡春鹤^1,2*, 谢将剑^1,2*

1. 北京林业大学工学院, 北京 100083; 2. 林木资源高效生产全国重点实验室, 北京 100083

收稿日期:2026-01-07 修回日期:2026-09-09 接受日期:2026-10-31
通讯作者: 胡春鹤, 谢将剑

Dense bird flock counting via feature enhancement and loss optimization

Guohui Liu^1,2, Chunhe Hu^1,2*, Jiangjian Xie^1,2*

1 School of Technology, Beijing Forestry University, Beijing 100083, China

2 State Key Laboratory of Efficient Production of Forest Resources, Beijing 100083, China

Received:2026-01-07 Revised:2026-09-09 Accepted:2026-10-31
Contact: Chunhe Hu, Jiangjian Xie
Supported by:
Supported by the National Key Research and Development Program of China(2024YFF1307202); the Beijing Natural Science Foundation(5252014); and Beijing Forestry University Science and Technology Innovation Program(2024XY-G002)

摘要/Abstract

摘要： 鸟群数量是评估生物多样性与种群动态的关键指标。基于摄像头的自动识别技术已成为野外候鸟监测的重要手段。然而, 从视频图像中实现精准计数仍面临三大挑战: 姿态多变与背景杂乱导致特征提取困难、远距成像引起目标尺度剧烈变化、密集重叠造成定位模糊与漏检。针对上述问题, 本文在聚焦逆距离变换(focal inverse distance transform, FIDT)定位图监督与HRNet (high-resolution network)高分辨率表征框架基础上, 构建面向密集鸟群计数任务的 DAHNet模型。该模型在HRNet主干中嵌入多尺度残差模块DC-basicblock增强特征表达; 在特征融合阶段引入融合通道注意力的联合金字塔上采样模块(JPU-SE)实现多尺度上下文信息自适应聚合; 构建由SmoothL1损失、I-SSIM损失和OT损失组成的复合损失函数, 协同提升FIDT定位图的生成质量。基于本文构建的密集鸟群图像数据集的实验结果表明, DAHNet的平均绝对误差(mean absolute error, MAE)和均方根误差(root mean square error, RMSE)分别为6.53和10.48, 相比HRNet + FIDT基线模型降低7.21和17.82, 并优于多种主流密度估计模型, 体现了DAHNet在本文数据集上的计数性能优势, 可为候鸟种群数量估计及相关生态监测提供技术支撑。

关键词: 鸟群计数, FIDT定位图, 多尺度特征融合, 生态监测, 损失优化

Abstract

Aims: Accurate counting of dense bird flocks from images remains challenging due to variations in bird posture, complex backgrounds, large differences in target scale, and occlusion. This study aims to improve bird counting accuracy in complex natural environments.

Method: We developed DAHNet based on focal inverse distance transform (FIDT) map supervision and the high-resolution network (HRNet). A multi-scale residual module (DC-basicblock) was incorporated to enhance feature extraction. A joint pyramid upsampling module with channel attention (JPU-SE) was introduced to aggregate multi-scale contextual information. Furthermore, a composite loss function combining SmoothL1, I-SSIM, and optimal transport (OT) loss was employed to improve the quality of predicted FIDT maps. Experiments were conducted on a dataset of 600 bird flock images.

Results: DAHNet achieved a mean absolute error (MAE) of 6.53 and a root mean square error (RMSE) of 10.48, representing absolute reductions of 7.21 and 17.82, respectively, compared with the HRNet + FIDT baseline. DAHNet also outperformed the selected comparison models on the test dataset.

Conclusion: DAHNet improves bird counting accuracy on the constructed dataset through feature enhancement, multi-scale fusion, and loss optimization. The proposed method may provide technical support for image-based bird population monitoring in complex natural environments.

Key words: bird flock counting, FIDT localization map, Multi-scale feature fusion, ecological monitoring, loss optimization.

刘国辉, 胡春鹤, 谢将剑. 基于特征增强与损失优化的密集鸟群计数算法. 生物多样性, DOI: 10.17520/biods.2026005.

Guohui Liu, Chunhe Hu, Jiangjian Xie. Dense bird flock counting via feature enhancement and loss optimization. Biodiversity Science, DOI: 10.17520/biods.2026005.

导出引用管理器 EndNote|Ris|BibTeX

链接本文: https://www.biodiversity-science.net/CN/10.17520/biods.2026005