Spectrogram-frame linear network and continuous frame sequence for bird sound classification

Here are the official TensorFlow implementation. This study is in the journal Ecological Informatics. You can find it here https://doi.org/10.1016/j.ecoinf.2019.101009

Highlights

1. Continuous frame sequence as a dynamic visual representation of the bird sound;
2. Spectrogram-frame linear network (SFLN) is used for the classification of bird sound, which emphasizes the changes of bird sound in time domain and frequency domain;
3. The shorter the frame length and frame shift of consecutive frame sequences, and the better the classification effect of SFLN;
4. The method achieves excellent classification effects and has good robustness in complex backgrounds.

Please cite if you referred our work.

Zhang X, Chen A, Zhou G, et al. Spectrogram-frame linear network and continuous frame sequence for bird sound classification[J]. Ecological Informatics, 2019, 54: 101009.

Apache License Version 2.0, January 2004 http://www.apache.org/licenses/

项目内容包括：

一个从 www.xeno-canto.org 下载音频数据的网络爬虫

• 下载列表维护，支持断点续传；
• 捕获网络异常，加入重试机制；
• 可根据重试计数更换代理；
• 模块化设计，可根据文件计数器加入动态代理；

一系列音频数据预处理过程

• 从MP3到梅尔频谱图的转换
• 图像标准化，数据增强
• continuous frame sequence 的构建
• 自动化的 batch 生成器

一个音频识别神经网络

• 以上述3D图像帧序列为输入
• 为音频特殊设计的 Spectrogram-frame linear layer
• 一系列结果分析以及可视化方法

环境配置：

• Tensorflow 1.14以上，需要eager模式支持
• librosa，音频处理
• PIL，图像处理
• scipy，图像处理
• matplotlib，绘图
• sklearn，需要用到一些结果分析函数