Automatic Insect Sound Recognition System Based On Convolutional Neural Network

Insect plays an important role in agricultural industry,whose species and quantity directly influence agricultural production and economic benefit.The solution to agricultural insect is still spraying pesticide,which indicates that the selection of the type and amount of pesticide is rather significant.As development of artificial intelligence,automatic insect recognition has become an outstanding achievement in agriculture domain.The two aspects of agricultural insect recognition technology are image recognition and sound recognition.In insect sound recognition area,traditional artificial sound feature extraction method is still widely used.However,this method inevitably lead to extra error because of researchers' subjective factors,which is caused by difficulties of feature selection and feature combination.This paper has studied important methods of insect sound feature extraction and then choose spectrogram as the original feature,whose R-space gray image is enhanced by contrast-limit adaptive histogram equalization(CLAHE).MFCC and chromatic spectrogram are used as input features,which is the contrast experiment.After insect sound recognition progress,recognition accuracy rate using enhanced spectrogram is far higher than that using MFCC or chromatic spectrogram.Considering of data scale,enhanced spectrogram has shown significant advantage in data dimension and data size.Then we have done a lot of research in classification technology.Traditional back propagation(BP)neural network and support vector machine(SVM)are most used in sound recognition area,whose efficiency is still low in large-scale data situation.Based on artificial intelligence and machine learning,we use convolutional neural network as classification technology and setup a CNN model containing two convolution layers,two pooling layers,two full-connection layers and an rectified linear unit(ReLU)layer.We use MFCC,chromatic spectrogram and enhanced spectrogram as input feature image and use BP neural network,SVM and our convolutional neural network model as classification method to execute experiments.Convolutional neural network has shown powerful advantage in training speed and recognition accuracy.After several contrast experiments,we eventually use enhanced spectrogram and convolutional neural network as our recognition method.We see 0.0798553 net error and 97.8723%recognition accuracy rate using enhanced spectrogram and CNN,compared with 0.296177 net error and 93.617%recognition accuracy rate using MFCC and CNN or 0.188119 net error and 95.7447%recognition accuracy rate using chromatic spectrogram.The data dimension of enhanced spectrogram is one-third of that of chromatic spectrogram.The data size 256×256 of enhanced spectrogram is far less than MFCC,whose data size is 3000×28.