Animal Sounds Detection And Recognition Based On Multilayer Energy Detection

With the development of economy, it leads to the destruction of the wildlife habitat and the population structure. And there is an increasing number of biological species threatened with extinction, so the work of the wildlife protection has become very urgent. However, the wild animals survive in general interpersonal rare deep forest or wilderness in reality, and people usually only hear their voice without reflecting their shape, which makes it difficult to study and protect them. In-depth research for the different animal sounds can help us better analyse and understand their survival status, it also helps to understand the ecological environment and forecast the changes in a particular region. But in the actual recording of the sound of the wild species, which is inevitably mixing with a variety of unpredictable environment noise, so researching the wild species sounds recognition in noisy situation has a more realistic value. Therefore, we study the wildlife species sounds recognition technology under different signal-to-noise ratios. The work consists mainly of the following:1) Signal detection:energy detection (ED) method of the traditional signal detection methods can not overcome the detection problem which is introduced by the uncertainty of the signal duration, so this paper uses a specific method of splitting the original observation vector continuously to construct the structure of the multilayer energy detection (MED) to select the most suitable length of the detector window, so that the window length can match the duration of the unknown sound signal. In addition, we only extract the modified sound features of the segments which pass through the detection of the energy detectors in the layer of the MED which have the best detection probability, so that we can select the segments that have stronger sound signal for training and testing.2) Feature extraction and selection:traditional Mel frequency cepstral coefficients (MFCC) has noise sensitivity, and Mel filters will make the spectral energy between the adjacent bands leak larger, so this paper extracts the features of the Gammatone filter cepstral coefficients(GFCC) which is more in line with the human auditory characteristics and the wavelet sub-band energy coefficient(WSEC) which is based on the Mel frequency scale, and then we modified the features for the wild animals sound recognition. Secondly, in order to take advantages of the respective advantages of the GFCC and WSEC features, this paper combines the two features as the final feature vector.3) Classification algorithms:for wild animal sound recognition, choosing the appropriate effective classifier related to the overall performance of the system. This paper takes advantage of the good stability of support vector machine (SVM) in noisy situations to construct the SVM model for the single features and the combination features which are extracted from the above respectively, and compares recognition performance under different signal-to-noise ratio through the experiment.This thesis regards birds as research object, studies the feature of different bird species songs to achieve the purpose of recognizing the different species, and the experimental results show that, the sound recognition system has a better noise robustness, and provides a theoretical and methodological basis for further analysis of the sound recognition of the different species of wild animals.