Research On The Classification Method Of Recyclable Garbage Based On Hyperspectral Image And Deep Learning

In recent years,the population has increased greatly,and the rhythm of people's life is getting faster and faster.This has led to the update of daily necessities and the emergence of a large number of various products in response to the demand.At the same time,it has led to the rapid increase in the output of household garbage and the diversification of garbage composition.This situation greatly increases the complexity of garbage classification and recycling.In this paper,based on the principle that the material characteristics differences of different recyclable garbage can be directly mapped to the spectral characteristics differences,the spectral information and spatial information of garbage can be obtained by hyperspectral imaging technology,and the classification of recyclable garbage data can be realized by combining the classification method of convolutional neural network.Building a push-broom hyperspectral imaging system in a laboratory optical darkroom environment to collect hyperspectral image information in the near infrared band of 799.513~1000.99 nm for five common recyclable garbage samples of paper,plastic,metal,glass and textile at an average 2.5nm spectral resolution,for avoiding the influence of different color information of the same type of sample on the spectral absorption characteristics in the visible light band.After the pretreatment of reflectivity correction,denoising and image background segmentation,the data of the the sample hyperspectral image was extracted in ROI.The principal component analysis algorithm was used to process the data extracted from the five recyclable garbage,and the feature space L = [802.063,807.164,814.815,819.916,827.566,873.472,891.325,919.378,947.432,955.083,995.888]nm was extracted.By designing and optimizing the classification model of support vector machine(SVM),the classification accuracy of the sample data set reached above 85%.The results show that the spectral information of different types of garbage in the feature space has good inter-class difference and intra-class similarity,which can replace the original data for classification research.Finally,the spectrum of the pixel neighborhood in the sample ROI is randomly extracted and reconstructed into a gray image to make up a database of 9450 numbers.Through the CNN analysis and design the 7 layer neural network model of CNN classification,select Stochastic Gradient Descent Method combined with periodic decay learning rate for CNN classification model,updating and optimizing each weighting parameters,the classification accuracy is more than 92% of the neural network.This paper verifies the scientific and feasibility of fast classification of recyclable garbage by hyperspectral imaging technology,and provides theoretical basis and practical guidance for solving the problem of garbage classification intelligently by hyperspectral imaging technology in the future.