Algorithms for classification of objects in images of the receiving camera of a modern reverse vending machines

Abstract

The work examines the problem of developing mathematical and software for image analysis in modern reverse vending machines. As part of the approach based on the use of shallow machine learning methods, the problem of processing images of objects subject to raw material processing in reverse vending machines is formulated as a task of segmenting these images, followed by subsequent classification by shape. A review and comparison of known segmentation methods is carried out in order to extract the shape of objects and generate features for classification. As a result of the comparison, a choice is made in favor of the active contours method. A relatively simple algorithm for classifying segmented objects based on “random forest” is proposed and studied. We also propose two algorithms for extracting classification features: an algorithm based on the analysis of the degree of fullness of parts of a segmented image and an algorithm that calculates the properties of the object area. As an alternative approach, a method for classifying the resulting images as a whole, without preliminary feature selection, based on deep learning, is also described. The problem of shortage of training data is discussed, and possible approaches to solving it are given. We describe the implementation of a convolutional neural network — a classifier with the DenseNet architecture, obtained using the transfer learning technique. The results of experiments to evaluate the effectiveness of the considered algorithms, conducted on the provided training set, are presented. The selected segmentation method, as well as both considered classifiers, demonstrated a high level of efficiency. When comparing the results of classification algorithms based on shallow (“random forest”) and deep machine learning (convolutional neural network), a choice was made in favor of the neural network approach when certain conditions were met for the training data.