Improving the quality of object classification in images by ensemble classifiers with stacking

Abstract

The object of research is the process of classifying objects in images. The quality of classification refers to the ratio of correctly recognized objects to the number of images. One of the options for improving the quality of classification is to increase the depth of neural networks used. The main difficulties along the way are the difficulty of training such neural networks and a large amount of computing that makes it difficult to use them on conventional computers in real time. An alternative way to improve the quality of classification is to increase the width of the neural networks used, by constructing ensemble classifiers with staking. However, they require the use of classifiers at the first stage with different structured processing of input images, characterized by high quality classification and relatively low volume of calculations. The number of known such architectures is limited. Therefore, the problem arises of increasing the number of classifiers at the first stage of the ensemble classifier by modifying known architectures. It is proposed to use blocks of rotation of images at different angles relative to the center of the image. It is shown that as a result of structured image processing by the starting classifier, processing of rotated image leads to redistribution of errors on image set. This effect makes it possible to increase the number of classifiers in the first stage of the ensemble classifier. Numerical experiments have shown that adding two analogs of the MLP-Mixer algorithm to known configurations of ensemble classifiers reduced the error from 1 to 11 % when working with the CIFAR-10 dataset. Similarly, for CCT, the error reduction was between 2.1 and 10 %. In addition, it has been shown that increasing the MLP-Mixer configuration in width gives better results than increasing in depth. A prerequisite for the success of using the proposed approach in practice is the structured image processing by the starting classifier

Reducing the volume of computations when building analogs of neural networks for the first stage of an ensemble classifier with stacking

“…The generalized architecture of the ensemble classifier is shown in Fig. 2 [9]. It consists of two stages.…”

“…Most of the considered approaches to constructing new architectures involve performing a certain volume of floating-point operations, which will lead not to a decrease but to an increase in the volume of calculations when designing analogs of neural networks. The simplest and most studied solution was proposed in [9]. Analogs are obtained by rotating the input images at different angles.…”

“…Convolutional neural networks show good results [10,11], but their processing is not structured; the entire image is processed at once. Therefore, constructing analogs based on them for use in the first stage of an ensemble classifier does not lead to an improvement in the quality of classification [9]. Transformer-based architectures were developed later than convolutional neural networks, are characterized by a high degree of structure and show better results [5].…”

“…This places a limit on the resulting improvement in classification quality of the entire ensemble classifier. To increase the number of neural networks at the first stage, it is necessary either to develop neural networks with a new processing architecture, or to build analogs of already known networks [9].…”

“…Such analogs should have the same processing structure as the basic neural networks but make errors on different sets of images. Work [9] describes one of the methods for constructing analogs by rotating the original images. Due to the structured nature of image processing in the original neural network, the analog makes classification errors on another set of images.…”