A NEURAL NETWORK CLASSIFIER FOR IDENTIFYING THE AUTHORSHIP OF WORKS CREATED BY IMPRESSIONIST ARTISTS
Authors:
1.
Kazanskiy nacional'nyy issledovatel'skiy tehnicheskiy universitet imeni A.N. Tupoleva
2.
Kazan National Research Technical University named after A.N. Tupolev
Type:
Article
Pages:
from 130
to 135
Status:
Published
Received:
06.02.2026
Accepted:
06.03.2026
Published:
05.04.2026
Language:
Russian
Keywords:
IMAGE CLASSIFICATION, AUTHORSHIP DETERMINATION, IMPRESSIONISM, NEURAL NETWORK MODELING, CONVOLUTIONAL NEURAL NETWORK, RESNET-50
Abstract:
This article addresses the problem of identifying the authorship of works by impressionist artists based on digital images, using the construction and analysis of a convolutional neural network model. The relevance of this work stems not only from the need to digitalize cultural heritage sites but also from the rapid development of generative neural networks, which pose new challenges in distinguishing graphic content created by humans and artificial intelligence. Impressionism, as one of the key artistic movements, is characterized by unique, yet difficult to formalize, features. Identifying the authorship of a work when all artists work within the same style is a very complex task, as differences in the author's brushstroke technique, color palette, and composition can be very subtle and insignificant. Therefore, modern artificial intelligence methods based on deep learning are appropriate for solving this problem. This paper employs an approach based on convolutional neural networks using transfer learning technology. The following main stages were completed: obtaining and preparing data for analysis, selecting libraries and architecture for constructing a neural network model, training the model, and evaluating its performance. The first stage was implemented using the publicly available Impressionist Classifier Dataset, obtained from the Kaggle platform and containing works by ten famous impressionist artists. The total number of images for classification was 4,978, of which the training set contained 3,990 images, and the validation set contained 988 images. The Python programming language, as well as the following libraries and frameworks, were used to build the convolutional neural network model: PyTorch, torchvision, NumPy, Pandas, Matplotlib, and others. The architecture of the ResNet-50 convolutional neural network, pretrained on the ImageNet database, was considered and implemented, adapted to solve the classification problem for 10 classes. A training strategy with frozen layers and subsequent fine-tuning was applied. The constructed (retrained) model demonstrated high efficiency, achieving 86.06% accuracy on the validation data set. An analysis of the model's classification quality was conducted using a confusion matrix, ROC curves, and other metrics (accuracy, Top-3 accuracy, precision, recall, and F1 score). Issues with classifying stylistically similar artists were identified, indicating the need for additional sample balancing and the use of specialized augmentation methods. Directions for future research are outlined.
This article addresses the problem of identifying the authorship of works by impressionist artists based on digital images, using the construction and analysis of a convolutional neural network model. The relevance of this work stems not only from the need to digitalize cultural heritage sites but also from the rapid development of generative neural networks, which pose new challenges in distinguishing graphic content created by humans and artificial intelligence. Impressionism, as one of the key artistic movements, is characterized by unique, yet difficult to formalize, features. Identifying the authorship of a work when all artists work within the same style is a very complex task, as differences in the author's brushstroke technique, color palette, and composition can be very subtle and insignificant. Therefore, modern artificial intelligence methods based on deep learning are appropriate for solving this problem. This paper employs an approach based on convolutional neural networks using transfer learning technology. The following main stages were completed: obtaining and preparing data for analysis, selecting libraries and architecture for constructing a neural network model, training the model, and evaluating its performance. The first stage was implemented using the publicly available Impressionist Classifier Dataset, obtained from the Kaggle platform and containing works by ten famous impressionist artists. The total number of images for classification was 4,978, of which the training set contained 3,990 images, and the validation set contained 988 images. The Python programming language, as well as the following libraries and frameworks, were used to build the convolutional neural network model: PyTorch, torchvision, NumPy, Pandas, Matplotlib, and others. The architecture of the ResNet-50 convolutional neural network, pretrained on the ImageNet database, was considered and implemented, adapted to solve the classification problem for 10 classes. A training strategy with frozen layers and subsequent fine-tuning was applied. The constructed (retrained) model demonstrated high efficiency, achieving 86.06% accuracy on the validation data set. An analysis of the model's classification quality was conducted using a confusion matrix, ROC curves, and other metrics (accuracy, Top-3 accuracy, precision, recall, and F1 score). Issues with classifying stylistically similar artists were identified, indicating the need for additional sample balancing and the use of specialized augmentation methods. Directions for future research are outlined.
Keywords:
IMAGE CLASSIFICATION, AUTHORSHIP DETERMINATION, IMPRESSIONISM, NEURAL NETWORK MODELING, CONVOLUTIONAL NEURAL NETWORK, RESNET-50
IMAGE CLASSIFICATION, AUTHORSHIP DETERMINATION, IMPRESSIONISM, NEURAL NETWORK MODELING, CONVOLUTIONAL NEURAL NETWORK, RESNET-50
