Bird Sounds Recognition Using Time-frequency Texture And Random Forests

Birds are not only the best friends of people, but one of the important members of the ecological environment, and the protection of birds plays an important role in maintaining the balance of natural ecological environment. While compared with the morphological characteristics, the sounds of birds are easy to record and more species specificity. Therefore, the identification of bird species by means of bird sounds recognition has an important practical significance. In the actual bird sounds recognition, bird sounds collected are often mixed with various unavoidable environmental noises. Thus considering the automatic recognition technology of bird sounds in a noisy environment has a better practical value. In light of this, in this paper we study an anti-noise recognition method of bird sounds based on time-frequency texture features and Random Forest classifier. The main work includes the followings:1) Front-end processing. As the traditional noise estimation algorithms usually require prior assumption that the background noise is steady and thus can not be used to deal with the real non-stationary environment noises, we propose a sound enhancement algorithm for the highly non-stationary environment by combining a dynamic noise estimation algorithm with the classical short-time spectrum audio enhancement technology, to realize the noise reduction of bird sounds.2) Syllable segmentation. After analyzing the variability, diversity as well as grammatical structure of bird sounds, we put forward a syllable segmentation algorithm based on the two-dimension time-frequency matrix. The algorithm can locate the start and end points of each syllable in a consecutive bird sound clip by comparing the amplitude spectrum of each time-frequency bin.3) Texture feature extraction. By observing the time-frequency graph of different bird sounds, we find they contain distinct texture pattern and have a very good distinction. So we introduce the texture analysis into the feature extraction of bird sounds and consider each segmented syllable as the basic texture unit. We extract five kinds of texture features by using the texture analysis theory of image processing on each syllable’s time-frequency graph. As calculating texture features with the traditional are not only computational overloading but serious waste of storage space, we take into account to use vectors to store the information of each co-occurrence gray pair in the Gray Level Co-occurrence Matrix, and the corresponding texture features are calculated by the obtained sum vector and difference vector.4) Recognition using the classifier. In view of the advantages of ensemble classifier in effectively improving the recognition accuracy and avoiding the overfitting of data, an ensemble classifier model based on decision tree-Random Forest is used for training and recognizing the texture feature vector of each syllable. And the final identification result of each test sample depends on the majority votings of all syllables which contained.The experimental results show that the algorithm proposed in this paper can recognize bird sounds accurately and quickly, better adapt to the different types of environmental noise and provide a reference for the further study of other types of ecological environmental sounds.