Research On Classification And Recognition Of Bird Songs Based On Deep Learning And Multi-dimensional Feature Fusion

Sound is a biologically important ecological feature.Understanding the characteristics of sound is of great significance to speech recognition and other fields.At present,the research on speech recognition based on human voice is very rich,but there are few researches focusing on bird song recognition.The research results related to bird song recognition play a great role in the collection of bird species information,bird behavior analysis,and ecological environment detection in nature reserves.Especially for the field of information collection in nature reserves,traditional information collection is mostly manual,it takes time and effort to determine the species of birds by observing the characteristics of birds.If the species of birds can be identified through the sound of birds,it will save a lot of work.It can be seen that the classification and recognition research based on bird song is very important.This paper proposes a bird song classification algorithm based on deep learning and multi-dimensional feature fusion.Through song spectrogram and feature extraction,combined with a deep learning model,bird species recognition can be realized through bird song.The main work and features of this thesis include:1.Use different data sets to establish a variety of sample libraries of birdsong spectrograms.In this paper,we construct a self-built dataset based on 26 common bird species in Chongming Dongtan and its surrounding areas,and uses two other public datasets to form the bird song dataset of this article.First,preprocesses the birdsong signal to obtain higher quality signal,and then uses the time-frequency analysis method to generate three types of spectrogram: STFT,Mel and Chirplet,each type of sample library including 60967 spectrograms.The sample library of spectrograms is the basis for the study of bird song recognition in this article.2.Designed a bird song classification method based on Res Net50-CNN-GRU.This paper compares the feature extraction capabilities of different migration models and finds that the Res Net50 migration model has unique advantages in bird song feature extraction,and then uses the gated recurrent unit(GRU)to process serialized data,combined with convolutional neural network(CNN)designed a bird song classification model based on CNN-GRU.Based on this,this paper has formed a bird song classification method based on Res Net50-CNN-GRU.3.This paper proposes a bird song recognition method based on multi-dimensional feature fusion.Different spectrogram can reflect different characteristics of bird song,and the best features of different spectrogram are merged to obtain fusion features.Experiments show that the fusion feature has better classification performance than a single feature,and can be verified on different migration models and data sets,it shows that our method is valid to improving bird song recognition effect.The experimental results show that the recognition algorithm proposed in this paper has a better classification effect than traditional bird song classification methods,and the improvement of the classification effect of the multi-dimensional feature fusion algorithm proposed in this paper can be verified in a variety of migration models.