基于特征迁移的多物种鸟声识别方法

Recognition of Multiple Bird Species in Audio Recordings Based on Feature Transfer

针对多物种鸟声识别中多物种鸟声样本不足的问题,尝试采用单物种鸟声样本训练多物种鸟声识别模型,并提出一种基于特征迁移的多物种鸟声识别方法。该方法引入特征迁移学习算法,利用最大均值差异(Maximum mean discrepancy,MMD)度量鸟声样本特征分布差异,将不同分布的单物种鸟声和多物种鸟声的音频特征映射为同分布的潜在音频特征,再基于同分布的音频特征构造识别模型。使得单物种鸟声样本训练的识别模型也能够适用于多物种鸟声识别。在自然形成的多物种鸟声数据集上,算法在4项多标记评价指标上都取得了较好的识别效果;在人工构造的多物种鸟声数据集上对比试验表明,基于特征迁移的识别算法在单个物种上的正确识别率相较于对比算法最高提升了20%。

To deal with the problem of inadequate sample in multiple bird species recognition, a new rec- ognition method of multiple bird species in audio recordings is proposed based on feature transfer, which uses bird sounds of single species to train a multiple bird species recognition model. Maximum mean dis- crepancy （MMD） is used to measure the feature distributions difference of bird sounds, which maps audio feature of single-species bird sounds and multiple-species bird sounds into a new latent feature with the same distribution. Then single-species bird sounds with latent feature can be used to train a model of multiple-species bird sounds. The experimental result shows that method can achieve good regognition performance in a natural multiple-species bird sounds dataset based on four multi-label metrics. The recognition rate of proposed method increases by 20 % compared with other methods in an artificial multiplespecies bird sounds dataset.