Page 100 - 《应用声学》2025年第2期

360
造的轻量级 EfficientNet-B0-Tuning 网络可以在资
源有限的情况下，在边缘设备上实现快速的鸟声识
别，为在生物多样性监测设备上部署声学识别模型
提供了技术支持。
表 8 在龙芯 2K1000LA 上平均推理时间
Table 8 Average inference time on
2K1000LA
模型 参数量/M
EfficientNet-B0 5.3
EfficientNet-B0-Tuning1 3.0
EfficientNet-B0-Tuning2 2.2
EfficientNet-B0-Tuning3 1.5
4 结论
针对被动鸟声检测的应用场景，特别是边缘设
备上的需求，本研究提出了一种轻量级鸟声识别方
法。首先，在网络结构化剪枝技术基础上，提出一种
基于逆背包准则进行 EfficientNet-B0 网络的逐步
剪枝方法。在推理上通过二级泰勒展开使用训练损
失来近似替代精度损失，再从损失中得到校准系数，
从而较快完成通道掩码剪枝。然后，提出一种加入
内部蒸馏损失分量的微调方法，将原基础网络知识
转移到剪切后模型中。其中，内部蒸馏损失分量通
过计算 MBConv 模块间的损失变化，在微调中达到
衡量模型全局损失的目的。实验表明，所提出的方
法不仅能够保持高识别精度，还能显著降低模型的
大小和计算复杂度，使其更适合资源受限的环境。
在权衡计算效率与精度上，本文实验中发现，选
择对复杂度高的深层网络进行压缩，要比直接选择
简单网络可以获得更好的识别精度。与传统基准网
络相比较，本文方法得到的轻量级网络在相同规模
条件下，在鸟声时频图分类精度上展示了其优越性，
可推广至其他声音事件识别的应用领域。但值得注
意的是，本文方法在获得了通道稀疏度较高、结构
稳定的网络结构的同时，网络调整的计算量仍较大，
这是未来研究的一个方向。
参 考 文 献
[1] Kahl S, Wood C M, Eibl M, et al. BirdNET: A deep learn- Applied Acoustics, 2022, 188: 108550.
ing solution for avian diversity monitoring[J]. Ecological [16] Tan M, Le Q V. EfficientNet: Rethinking model scaling
Informatics, 2021, 61(7): 101236.
2025 年 3 月
[2] Adavanne S, Drossos K, Çakir E, et al. Convolutional re-
current neural networks for bird audio detection[C]// 2017
25th European Signal Processing Conference (EUSIPCO),
Kos, GR, 2017: 1744–1748.
[3] Kahl S, Wilhelm-Stein T, Hussein H, et al. Large-scale
bird sound classification using convolutional neural net-
works[C]// Conference and Labs of the Evaluation Forum
2017. Amsterdam, NL, 2017: 1–14.
[4] Zhang C, Jiang W, Zhao Q, et al. Semantic segmentation
of aerial imagery via split-attention networks with disen-
推理时间/s tangled nonlocal and edge supervision[J]. Remote Sensing,
0.41 2021, 13(6): 1176.
0.24 [5] Brock A, De S, Smith S L, et al. High-performance large-
scale image recognition without normalization[C]// Pro-
0.19
ceedings of the 38th International Conference on Machine
0.13 Learning (PMLR), 2021, 139: 1059–1071.
[6] Xie S, Girshick R, Dollar P, et al. Aggregated residual
transformations for deep neural networks[C]// 2017 IEEE
Conference on Computer Vision and Pattern Recognition
(CVPR), Honolulu, HI, USA, 2017: 5987–5995.
[7] Liu Z, Mao H, Wu C Y, et al. A ConvNet for the
2020s[C]// 2022 IEEE Conference on Computer Vision
and Pattern Recognition (CVPR), arXiv: 2201.03545,
2022.
[8] Wang Y, Li S X, Zhang H, et al. A lightweight CNN-
based model for early warning in sow oestrus sound mon-
itoring[J]. Ecological Informatics, 2022, 72: 101863.
[9] Wang Y, Zhou J G, Zhang C Y, et al. Bird object detec-
tion: dataset construction, model performance evaluation,
and model lightweighting[J]. Animals, 2023, 13(18): 2924.
[10] Incze A, Jancso H B, Szilagyi Z, et al. Bird sound recogni-
tion using a convolutional neural network[C]// 2018 IEEE
16th International Symposium on Intelligent Systems and
Informatics (SISY), Subotica, SA, 2018: 295–300.
[11] Hu S P, Chu Y H, Tang L, et al. A lightweight multi-
sensory field-based dual-feature fusion residual network
for bird song recognition[J]. Applied Soft Computing,
2023, 146: 110678.
[12] Molchanov P, Mallya A, Tyree S, et al. Importance esti-
mation for neural network pruning[C]//2019 IEEE/CVF
Conference on Computer Vision and Pattern Recognition
(CVPR), Long Beach, CA, USA, 2019: 11256–11264.
[13] Bai H, Wu J, King I, et al. Few shot network compression
via cross distillation[C]// Proceedings of the AAAI Con-
ference on Artificial Intelligence. New York, USA, 2020:
3203–3210.
[14] Aguilar G, Ling Y, Zhang Y, et al. Knowledge distilla-
tion from internal representations[C]// Proceedings of the
AAAI Conference on Artificial Intelligence. New York,
USA, 2020: 7350–7357.
[15] Xie J, Zhao S, Li X, et al. KD-CLDNN: Lightweight au-
tomatic recognition model based on bird vocalization[J].
for convolutional neural networks[J]. arXiv: 1905.11946,