The number of sneezing increase as swine influenza infection symptom in an early stage. Collecting many sneezing sounds of infected pigs is hard; sneezing classifier used small size acoustic features is necessary. In previous research, F-measure (of classifying accuracy) was about 60 % only, moreover, comparative evaluation has not conducted in a different environment and different acoustic features. The purpose of this paper is developing a pig sneezing classifier detectable in a different recording environment on high performance. We recorded a video and acoustic signal in multiple positions for 2 weeks after we infected pigs with swine influenza. In the experiment, we used multiple kinds of influenza virus. From the recorded acoustic signal, we detected 74533 samples of acoustic events automatically under a decided detection level. We assigned labels using with a movie for a part of acoustic events; we collected acoustic events including 144 sneezes. For acoustic events, we extracted a variety of acoustic features, and we evaluated classification performance using a classifier based on Support Vector Machine. As a result, developed classifier’s F-measure is 92.8 %, and it is very higher than the previous method. In this case, the classifier’s acoustic features are Mel Frequency Cepstral Coefficients, a feature explained spectral rising, and frequency change in a low-frequency band. In addition, trained classifier detected 3764 sneezes. Consequently, we developed high-performance sneezing classifier using small size acoustic features for detectable in a different recording environment.
View full abstract