An open-set domain adaptation method for wildlife image recognition via adversarial disentanglement and feature alignment

doi:10.17520/biods.2025283

Abstract

Abstract:

Aim: Wildlife is a vital component of biodiversity, and its efficient monitoring through image recognition is crucial for conservation. However, the performance of wildlife image recognition models often declines due to cross-domain distribution shifts from complex environments and interference from unknown species in the target domain.

Methods: To address these challenges, we propose an open-set domain adaptation method for wildlife images that integrates adversarial disentanglement and feature alignment. We first constructed a domain adversarial network using the ResNet50 residual network. Next, a dual optimization strategy combining center alignment and orthogonal projection was employed to enhance the discriminative power for known categories and disentangle the feature space of unknown categories. The final open-set domain adaptation recognition model was developed by integrating these components.

Results: When trained and evaluated on datasets with 8 and 11 wildlife species, our method achieved Average-HOS values of 48.95% and 46.38%, respectively. This represents a significant performance improvement of 10.90% and 5.37% in Average-HOS compared to the best baseline methods.

Conclusion: The collaborative optimization approach effectively addresses domain shift and unknown category interference, thereby enhancing the model’s cross-domain generalization and unknown category identification capabilities in real-world scenarios.

Key words: wildlife, image recognition, open-set domain adaptation, orthogonal projection loss, center loss

Jianing An, Changchun Zhang, Jiantao Wang, Zhiyong Pei, Dandan Bai, Junguo Zhang. An open-set domain adaptation method for wildlife image recognition via adversarial disentanglement and feature alignment[J]. Biodiv Sci, 2025, 33(12): 25283.

Add to citation manager EndNote|Ris|BibTeX

URL: https://www.biodiversity-science.net/EN/10.17520/biods.2025283

https://www.biodiversity-science.net/EN/Y2025/V33/I12/25283

Figures/Tables 15

References 29

[1]	Bai SH, Zhang M, Zhou WQ, Huang ST, Luan ZR, Wang DL, Chen BD (2024) Prompt-based distribution alignment for unsupervised domain adaptation. In: 2024 AAAI Conference on Artificial Intelligence, pp.729-737. AAAI Press, Vancouver.
[2]	Bucci S, Loghmani MR, Tommasi T (2020) On the effectiveness of image rotation for open set domain adaptation. In: 2020 European Conference on Computer Vision (ECCV), pp.422-438. Springer International Publishing, Glasgow.
[3]	Busto PP, Gall J (2017) Open set domain adaptation. In: 2017 IEEE International Conference on Computer Vision (ICCV), pp. 754-763. IEEE, Venice.
[4]	Chang DL, Sain A, Ma ZY, Song YZ, Wang RP, Guo J (2024) Mind the gap: Open set domain adaptation via mutual-to-separate framework. IEEE Transactions on Circuits and Systems for Video Technology, 34, 4159-4174. DOI URL
[5]	Ganin Y, Lempitsky V (2015) Unsupervised domain adaptation by backpropagation.In: 2015 IEEE International Conference on Machine Learning (ICML), pp. 1180-1189. IEEE, Lille.
[6]	Gao F, Pi DC, Chen JF (2024) Balanced and robust unsupervised open set domain adaptation via joint adversarial alignment and unknown class isolation. Expert Systems with Applications, 238, 122127. DOI URL
[7]	He KM, Zhang XY, Ren SQ, Sun J (2016) Deep residual learning for image recognition. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 770-778. IEEE, Las Vegas.
[8]	He ZW, Zhang ZL, Zhang L (2024) Survey on cross-domain object detection in open environment. Journal of Computer-Aided Design & Computer Graphics, 36, 485-502. (in Chinese with English abstract)
	[何贞苇, 张治龙, 张磊 (2024) 开放环境下的跨域物体检测综述. 计算机辅助设计与图形学学报, 36, 485-502.]
[9]	Jing TT, Liu HF, Ding ZM (2021) Towards novel target discovery through open-set domain adaptation. In: 2021 IEEE International Conference on Computer Vision (ICCV), pp. 9302-9311. IEEE, Montreal.
[10]	Liu H, Cao ZJ, Long MH, Wang JM, Yang Q (2019) Separate to adapt: Open set domain adaptation via progressive separation. In: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 2922-2931. IEEE, Long Beach.
[11]	Loghmani MR, Vincze M, Tommasi T (2020) Positive-unlabeled learning for open set domain adaptation. Pattern Recognition Letters, 136, 198-204. DOI URL
[12]	Long SF, Wang SS, Zhao X, Fu ZH, Wang BL (2023) Sample separation and domain alignment complementary learning mechanism for open set domain adaptation. Applied Intelligence, 53, 18790-18805. DOI
[13]	Lu JY, Shi XY, Duo LA, Wang TM, Li ZL (2024) Circadian rhythms of urban terrestrial mammals in Tianjin based on camera trapping method. Biodiversity Science, 32, 23369. (in Chinese with English abstract) DOI
	[卢佳玉, 石小亿, 多立安, 王天明, 李治霖 (2024) 基于红外相机技术的天津城市地栖哺乳动物昼夜活动节律评价. 生物多样性, 32, 23369.] DOI
[14]	Ma ZB, Dong YQ, Xia Y, Xu DL, Xu F, Chen FX (2024) Wildlife real-time detection in complex forest scenes based on YOLOv5s deep learning network. Remote Sensing, 16, 1350. DOI URL
[15]	Qu SQ, Zou TP, He LH, Röhrbein F, Knoll A, Chen G, Jiang CJ (2024) LEAD: Learning decomposition for source-free universal domain adaptation.In: 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 23334-23343. IEEE, Seattle.
[16]	Ranasinghe K, Naseer M, Hayat M, Khan S, Khan FS (2021) Orthogonal projection loss.In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pp. 12313-12323. IEEE, Montreal.
[17]	Roy AM, Bhaduri J, Kumar T, Raj K (2023) WilDect-YOLO: An efficient and robust computer vision-based accurate object localization model for automated endangered wildlife detection. Ecological Informatics, 75, 101919. DOI URL
[18]	Saito K, Yamamoto S, Ushiku Y, Harada T (2018) Open set domain adaptation by backpropagation. In: 2018 European Conference on Computer Vision (ECCV), pp.156-171. Springer International Publishing, Munich.
[19]	Tabak MA, Norouzzadeh MS, Wolfson DW, Sweeney SJ, Vercauteren KC, Snow NP, Halseth JM, Salvo PA, Lewis JS, White MD, Teton B, Beasley JC, Schlichting PE, Boughton RK, Wight B, Newkirk ES, Ivan JS, Odell EA, Brook RK, Lukacs PM, Moeller AK, Mandeville EG, Clune J, Miller RS (2019) Machine learning to classify animal species in camera trap images: Applications in Ecology.Methods in Ecology and Evolution, 10, 585-590.
[20]	Tan SH, Jiao JN, Zheng WS (2019) Weakly supervised open-set domain adaptation by dual-domain collaboration.In: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 5389-5398. IEEE, Long Beach.
[21]	Wang DW, Cheng S, Feng JW, Wang TM (2025) The wildlife camera-trapping dataset of Zhangguangcai Mountains in Northeast China (2015-2020). Biodiversity Science, 33, 24384. (in Chinese with English abstract) DOI
	[王大伟, 程帅, 冯佳伟, 王天明 (2025) 东北地区张广才岭2015-2020年野生动物红外相机监测数据集. 生物多样性, 33, 24384.] DOI
[22]	Wen YD, Zhang KP, Li ZF, Qiao Y (2016) A discriminative feature learning approach for deep face recognition. In: 2016 European Conference on Computer Vision (ECCV), pp. 499-515. Springer International Publishing, Amsterdam.
[23]	Xiao ZS, Xiao WH, Wang TM, Li S, Lian XM, Song DZ, Deng XQ, Zhou QH (2022) Wildlife monitoring and research using camera-trapping technology across China: The current status and future issues. Biodiversity Science, 30, 22451. (in Chinese with English abstract) DOI
	[肖治术, 肖文宏, 王天明, 李晟, 连新明, 宋大昭, 邓雪琴, 周岐海 (2022) 中国野生动物红外相机监测与研究: 现状及未来. 生物多样性, 30, 22451.] DOI
[24]	Yousif H, Kays R, He Z (2019) Dynamic programming selection of object proposals for sequence-level animal species classification in the wild. IEEE Transactions on Circuits and Systems for Video Technology, 2, 28-29.
[25]	Zhang CC, Zhang JG (2023) DJAN: Deep joint adaptation network for wildlife image recognition. Animals, 13, 3333. DOI URL
[26]	Zhao ET, Zhang CC, Zhao HT, Zhang JG (2025) A recognition method of domain adaptation for wildlife images based on adversarial learning. Scientia Silvae Sinicae, 61(4), 1-8. (in Chinese with English abstract)
	[赵恩庭, 张长春, 赵海涛, 张军国 (2025) 基于对抗学习的野生动物图像域适应识别方法. 林业科学, 61(4), 1-8.]
[27]	Zheng F, Lü LX, Lü HX, Shi J, Li R, Yang SL, Liu BW (2025) The research on the mammal diversity and its dynamic based on camera-trapping in Heilongjiang Liangshui National Nature Reserve. Acta Ecologica Sinica, 45, 5289-5296. (in Chinese with English abstract)
	[郑霏, 吕来新, 吕泓学, 师杰, 李瑞, 杨世丽, 刘丙万 (2025) 基于红外相机监测黑龙江凉水国家级自然保护区兽类多样性及其变化. 生态学报, 45, 5289-5296.]
[28]	Zheng JP, Wen YB, Chen MX, Yuan S, Li WJ, Zhao Y, Wu WZ, Zhang LX, Dong RM, Fu HH (2024) Open-set domain adaptation for scene classification using multi-adversarial learning. ISPRS Journal of Photogrammetry and Remote Sensing, 208, 245-260. DOI URL
[29]	Zhong L, Fang Z, Liu F, Yuan B, Zhang GQ, Lu J (2023) Bridging the theoretical bound and deep algorithms for open set domain adaptation. IEEE Transactions on Neural Networks and Learning Systems, 34, 3859-3873. DOI URL

数据集 Datasets	野猪 Wild boar	中华斑羚 Chinese goral	马鹿 Wapiti	猞猁 Eurasian lynx	貉 Raccoon dog	狗獾 Eurasian badger	东北兔 Manchurian hare	狍 Eastern roe deer
D	911	305	315	79	193	214	25	168
N	1,449	758	229	136	560	429	401	25

数据集 Datasets	野猪 Wild boar	中华斑羚 Chinese goral	马鹿 Wapiti	猞猁 Eurasian lynx	貉 Raccoon dog	狗獾 Eurasian badger	东北兔 Manchurian hare	狍 Eastern roe deer
D	911	305	315	79	193	214	25	168
N	1,449	758	229	136	560	429	401	25

模型 Model	迁移任务 Transfer task D→N				迁移任务 Transfer task N→D				Average-HOS (%)
模型 Model	OS (%)	OS* (%)	UNK (%)	HOS (%)	OS (%)	OS* (%)	UNK (%)	HOS (%)	Average-HOS (%)
ResNet50	28.00	25.07	51.51	33.72	29.15	25.29	52.33	34.10	33.91
DANN	33.55	30.85	49.78	38.09	34.87	33.31	44.24	38.00	38.05
ROS	28.91	20.87	77.13	32.85	31.91	30.86	38.20	34.14	33.50
ROS*	24.36	20.58	47.05	28.63	30.25	28.89	38.42	32.98	30.81
MTS	31.78	37.08	25.78	30.41	22.15	25.84	39.26	31.17	30.79
本文方法 Ours	43.21	41.21	55.16	47.18	44.32	40.30	68.39	50.72	48.95

模型 Model	迁移任务 Transfer task D→N				迁移任务 Transfer task N→D				Average-HOS (%)
模型 Model	OS (%)	OS* (%)	UNK (%)	HOS (%)	OS (%)	OS* (%)	UNK (%)	HOS (%)	Average-HOS (%)
ResNet50	28.00	25.07	51.51	33.72	29.15	25.29	52.33	34.10	33.91
DANN	33.55	30.85	49.78	38.09	34.87	33.31	44.24	38.00	38.05
ROS	28.91	20.87	77.13	32.85	31.91	30.86	38.20	34.14	33.50
ROS*	24.36	20.58	47.05	28.63	30.25	28.89	38.42	32.98	30.81
MTS	31.78	37.08	25.78	30.41	22.15	25.84	39.26	31.17	30.79
本文方法 Ours	43.21	41.21	55.16	47.18	44.32	40.30	68.39	50.72	48.95

模型 Model	迁移任务 Transfer task S1→S2				迁移任务 Transfer task S2→S1				Average-HOS (%)
模型 Model	OS (%)	OS* (%)	UNK (%)	HOS (%)	OS (%)	OS* (%)	UNK (%)	HOS (%)	Average-HOS (%)
ResNet50	28.00	25.07	51.51	33.72	37.91	37.25	43.15	39.99	36.85
DANN	32.58	30.22	51.40	38.06	39.52	37.92	52.27	43.96	41.01
ROS	23.58	18.39	65.10	28.68	23.94	19.37	60.51	29.34	29.01
ROS*	16.58	8.05	84.82	14.70	20.24	12.57	81.61	21.78	18.24
MTS	27.08	25.94	36.20	30.22	35.83	34.58	33.42	33.99	32.11
本文方法 Ours	38.05	35.04	62.10	44.80	41.80	39.34	61.43	47.96	46.38