Two-Level Classification Method Based On Sound Time-Frequency Spectrum Image Feature

The research of the identification of environmental sounds is increasingly affecting all areas of society, at the same time, because there will be often a lot of background noise in the sound recognition under the actual environment, it is particularly important to study how to accurately identify meaningful sound information from background noise. Aiming to this problem, in this paper, we present a method in which the identification and classification of sound is completed by the analysis of the sound color time-frequency spectrum image. Using the step-by-step clustering algorithm, we primarily conduct the first level clustering of sound signals; Then, through the dual-threshold pseudo-color mapping algorithm based on Intensity Slicing and HSV color map, the time-frequency matrix of each cluster of sound signals is mapped to color image; Finally, the color moments feature of time-frequency spectrum image which get different weights is extracted, and then, to each cluster, the SVM is used to classify the color moments feature vectors, as the second level classification. In the algorithm proposed process, this paper mainly proposed three innovation:(1) Improving the Endpoint detection algorithm. Due to the periodicity of effective sound segment, basing on the traditional endpoint detection methods, we propose a two-grades differentiating endpoint detection algorithm.(2) Proposing the step-by-step clustering algorithm. Because KNN classification process requires the known training sample types to achieve the classification of the classified samples, at the same time, in order to solve the difficult selection problem of k value in KNN algorithm and avoid calculating the similarity distance of every classified samples and all samples, this paper according to the needs of the recognition algorithm proposes a step-by-step clustering algorithm based on KNN algorithm thinking. Experimental results show that this algorithm can be good at the first level clustering of audio signals.(3) Proposing the dual-threshold pseudo-color mapping algorithm based on Intensity Slicing and HSV color map. In order to transform the audio features analysis to the analysis of the image features, we propose the dual-threshold pseudo-color mapping algorithm that integrates the Intensity Slicing technology and HSV color mapping technology, which transforms the time-frequency matrix of the audio data into color time-frequency spectrum image by mapping. The algorithm can effectively transform the noise component and the effective sound component into different monochrome regions, in order to achieve the separation of noise.Through the proposed method in this paper, we have applied the image processing technology to the identification and classification of environmental sounds. And, experimental results show that under the noisy environment this method have good anti-noise performance and classification effect.