Audio Recognition Method Based On The Saliency Detection Of Spectrogram

The paper studies audio recognition problems under complex environment. According to the correlation of spectrogram and sound signal and the significant recognition detection method in image processing, this paper propose a new audio recognition method based on the saliency detection of spectrogram. The experimental results show that the audio recognition method based on the saliency detection of spectrogram can effectively solve the identification problem of complex audio sources. In this paper, the main research work and achievements are as follows:1. The paper comprehensively introduces several classical algorithms for significant detection and image feature extraction, it describes in detail Itti algorithm and GBVS algorithm which are related to spectrograms recognition. Through a detailed analysis and comparison, the paper illustrates the principle of all kinds of classical significance detection algorithm and feature extraction algorithm, and the theoretical basis and application scope.2. The paper proposes a new audio recognition method based on the saliency detection of spectrogram. This method convert audio recognition in speech signal processing to saliency detection in image processing through the correlation of spectrogram and sound signal, meanwhile realize the audio feature modeling recognition based on spectrogram by detecting and recognizing spectrogram. This algorithm has high general applicability and effectiveness.3. The paper proposes a novel main spectrogram separation method applied to spectrograms significance representation mechanism. It overcomes the defect of manual annotation in main sound source when analyzing sound source in spectrogram. At the same time, the paper adds the most effective characteristics in main sound source region to the main spectrogram separation method in order to reduce the interference to sound source by sound signals, such as additive noise, multiple sound source, sound source splicing or distortion signal, thus enhance the generalization ability of main image model.4. The paper designs a set of relative perfect experiment based on the spectrogram audio recognition algorithm, which use various type of experimental data to compare, analysis and research largely. It fully demonstrates that the main image separation method for spectrogram significant testing has robustness and effectiveness. Finally it verifies that the algorithm described in this paper has good anti-noise recognition effect for complex audio in the real environment. The algorithm has better recognition performance than traditional audio recognition.