基于智能手机声信号哈密瓜成熟度的快速检测

Fast Detection of Hami Melon Ripeness Based on Features Extracted from Acoustic Signals of Smartphone

为满足瓜农和消费者便携、快速、价廉无损检测哈密瓜成熟度的要求,采用手机录制不同成熟度哈密瓜的拍打声信号并进行分析处理,提取11 个特征量,然后选择对不同成熟度具有显著差异的单个或多个特征量组成不同特征向量训练支持向量机分类器,通过对哈密瓜的未熟、适熟和过熟3 种成熟度判别结果的混淆矩阵分析,确定采用频谱质心wc、帧能量E、第1子带短时能量比SSTE1组成特征向量训练成熟分类器,最适于哈密瓜未熟瓜和成熟瓜判别;采用过零率、第2子带短时能量比SSTE2、第3子带短时能量比SSTE3组成特征向量训练适熟分类器,最适于哈密瓜适熟瓜和过熟瓜判别。该研究开发的手机安卓应用程序对哈密瓜成熟度判别总体准确率可达90.9%,并可通过用户反馈进一步提高判别能力。与此同时,采用逐步多元回归预测模型,实现wc、E及第1、2、4子带短时能量对哈密瓜糖度较准确的预测。

This work aimed to address the desire of both melon growers and consumers for convenient, rapid, low-costand non-destructive detection of Hami melon ripeness. In this paper, a mobile phone was used to record acoustic signalsgenerated by thumping Hami melons with different maturity levels. Then the raw signals were preprocessed to remove noisesand segmented into thumping event frames. A series of extraction steps were performed on these acoustic signals, includingthe calculation of root mean square, start/end detection, extraction of thumping events, filtration with Butterworth filter andfast Fourier transform. As a result, eleven features including short-time energy (STE), frame energy (E), average amplitudedifference function (AMDF), short-time energy ratios of four sub-bands (SSTE1, SSTE2, SSTE3 and SSTE4), zero crossingrate (ZCR), spectral centroid (wc), resonant frequency (f ) and bandwidth (B) were extracted. The significant differencesamong these features were analyzed and the ripeness-related features were found. Eight features (i.e. STE, E, SSTE1, SSTE2,SSTE4, ZCR, wc and f ) were selected to construct a ripe classifier-based support vector machine (SVM) for distinguishingbetween unripe and ripe melons. Also, seven features (i.e. AMDF, SSTE1, SSTE2, SSTE3, SSTE4, ZCR, and B)were selected to construct a proper ripe SVM classifier for distinguishing properly ripe from overripe melons. For both SVMclassifiers, the algorithm used radial base kernel function to achieve better performance. Single or multiple selected ripenessrelatedfeatures were used as vector to train the SVM classifiers. We measured the recall, precision, accuracy and F1-Measureby confusion matrix analysis for the two classifiers. The results showed that the classifier trained by the feature vectorswc, E and SSTE1 was the most suitable for disguising unripe from ripe melons. While the classifier trained by the featurevectors ZCR, SSTE2 and SSTE3 was the most suitable to discriminate properly ripe from overripe melons. Furthermore,a prediction model for the soluble solids content of melon was constructed by stepwise multiple regressions with wc, E,SSTE1, SSTE2 and SSTE. Finally, we developed an Android application on the mobile phone, which can correctly classify the melon maturities with an overall accuracy of 90.9%. Its classification performance can be improved by the user feedback.Additionally, it can be used to quantitatively detect the sugar content of melon with a high prediction accuracy.