MODELING AND EVALUATION OF DEEP NEURAL NETWORK RESNET-50 FOR CLASSIFICA-TION OF HOUSEHOLD SOLID WASTE

Abstract

The article presents the results of a study of the effectiveness of the deep convolutional neural network ResNet-50, adapted for the task of automated classification of municipal solid waste. The aim of the work is to assess the accuracy of the model, identify typical classification errors and justify the possibilities of its integration into industrial sorting lines. The study used an extended dataset, which includes images of five main categories of waste: plastic, glass, metal, paper and organic materials. Pre-processing of the data involved the use of complex augmentation, which ensured the model's resistance to noise, changes in lighting, deformations and contamination of objects. To assess the effectiveness, the metrics of accuracy, completeness, F1-measure and error matrix were used, which allowed to establish structural patterns of false predictions. The results obtained demonstrate high classification quality for most classes, in particular, accuracy over 0.95, as well as stable agreement of training and validation curves without signs of overtraining. Analysis of the error matrix revealed a number of typical errors associated with the similarity of textural and spectral properties of individual classes, which is especially characteristic of plastic and organic waste, as well as paper and cardboard. Approaches to their elimination are proposed, in particular, the use of spectral data (NIR / MIR), the expansion of augmentations and fusion models of features. The practical significance of the study lies in the possibility of integrating the constructed model into robotic waste sorting systems, which ensures increased processing efficiency, reduced human factor and optimization of the operation of modern waste sorting complexes. The presented results can be used to create new generation industrial classifiers and develop intelligent waste management systems.