Multi-feature Detection Method And Experimental Research Of Construction And Demolition Waste Based On Hyperspectral Vision

The current processing technology of construction and demolition waste(CDW)is less intelligent,resulting in a low recovery rate.Intelligent sorting based on online detection technology can significantly improve processing efficiency.This paper combines 3D and hyperspectral vision to develop an intelligent detection system for CDW.According to different working conditions,two detection methods are studied.Solution(1)studies the machine learning method of pixel-level hyperspectral features,which can achieve an average F1 accuracy of 98.52%.The main research contents include: 1)The principal component analysis(PCA)and wavelet transform(WT)are used for dimensionality reduction to extract the amplitude features of the spectral curves.2)The first-order derivative(FD)and standard normal variate transformation(SNV)are used to extract trend features of the spectral curves.3)Using extreme learning machine(ELM)and random forest(RF)to identify the extracted features,it is found that RF has significant advantages in identifying trend features,while ELM is affected by the scale and distribution of input features,and the overall accuracy is slightly lower than RF.4)Due to the fluctuation of the light source in practical,the adaptability of every features to the fluctuation are studied,and it is found that although the SNV feature is more recognizable,it is most seriously affected by the fluctuation.5)Recognizing WT features through ELM combined with RF recognizing FD features are used as the online solution.Using 3D contours to eliminate background,segmenting the recognition results,and counting the proportion of each category within the target contour after segmentation.At last,calculating the centroid and pose based on the contour.In order to solve the problem that solution(1)cannot deal with the adhesion and stacking of similar objects,solution(2)based on instance segmentation is studied.The highest m AP accuracy is 96.10%,while ensuring generalization,through optimizing the training mode of Mask RCNN.The main research contents include: 1)Comparing the training process of hyperspectral images,spectral pseudocolor images,and color images,it is found that the main factors affecting the recognition accuracy include the initial state of the model and the relative data volume.2)Experimental verifications are carried out for the two influencing factors.Mutual promotion and mutual expansion between different data are proposed,which are used to improve the initial state of the model and the relative data volume respectively.The combination of the two methods can stably improve the performance of the model.The color image and the pseudo-color image can mutually carry out data expansion and promote training,which can reduce the labeling workload and cost in practical applications.Solution(1)is to study the characteristics of different features and find the suitable recognition algorithm,which improves the accuracy of the two types of features by0.48% and 2.15% respectively.However,due to the characteristics of its single-point processing method,it is impossible to deal with the problems of adhesion and stacking of different materials.Solution(2)is based on the data characteristics of hyperspectral images,and proposes a program to promote training and enhanced training,improve the generalization performance of the instance segmentation method,and have the function of dealing with adhesion and stacking problems.On the basis of the researches above,the CDW intelligent detection system developed in this paper can automatically detect the target location on the conveyor,and can segment and identify the target at the same time.Combined with mature robot technology,it can realize automatic sorting of CDW and improve processing efficiency.