NEURAL NETWORK CONVOLUTIONAL MODELS FOR DETERMINING HUMAN AGE FROM A PHOTOGRAPH
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 113
to 119
Status:
Published
Received:
31.01.2026
Accepted:
31.01.2026
Published:
31.01.2026
Language:
Russian
Keywords:
INTELLIGENT SYSTEM, HUMAN AGE, NEURAL NETWORK MODELING, CONVOLUTIONAL NEURAL NETWORK, YOLOV8N, RESNET50V2, IMAGE CLASSIFICATION
Abstract:
This article explores the problem of determining a person's age from a photograph using the construction and analysis of convolutional neural network models. It is shown that age is a significant factor in image analysis for a number of applied tasks. The application of automatic age determination systems is broad and includes such areas as video surveillance and access control systems, content filtering in online services, marketing and advertising personalization, digital technologies in education, medical diagnostics, and augmented and virtual reality. Age-related changes typically affect the following characteristics: wrinkles, facial contour changes, changes in skin color and texture, drooping eyelids, and changes in the shape of the lips and chin. Visual cues can be masked by cosmetics, medical procedures, lifestyle, and may also vary depending on ethnicity. This makes age determination from an image particularly challenging. Therefore, modern deep learning methods are appropriate for solving the age determination problem. This paper uses an approach based on convolutional neural networks to solve this problem. The following stages were completed for its implementation: obtaining and preparing data for analysis, selecting libraries and architectures for building convolutional neural network models, training convolutional neural network models, and evaluating the model performance. The following publicly available datasets were selected for the implementation of the first stage: Facial Age and UTKFace. A total of three data preparation scenarios were implemented: classification by exact age, classification by eleven age ranges, and classification by five coarse-grained age categories. The total number of images for classification was approximately 14,000, of which the training set contained 9,734 and the validation set contained 4,238 images. The Python programming language, as well as the TensorFlow, NumPy, Matplotlib, Seaborn, Scikit-learn, Pillow, and Tkinter libraries and frameworks, were used to build the neural network models. Two convolutional neural network architectures are considered and compared: ResNet50V2 and YOLOv8n. The YOLOv8n architecture demonstrated superiority over ResNet50V2 for each of the metrics: Accuracy, Macro F1-score, and Weighted F1-score. Issues with classifying middle-aged groups were identified, indicating the need for additional balancing of the training set and the use of augmentation methods.
This article explores the problem of determining a person's age from a photograph using the construction and analysis of convolutional neural network models. It is shown that age is a significant factor in image analysis for a number of applied tasks. The application of automatic age determination systems is broad and includes such areas as video surveillance and access control systems, content filtering in online services, marketing and advertising personalization, digital technologies in education, medical diagnostics, and augmented and virtual reality. Age-related changes typically affect the following characteristics: wrinkles, facial contour changes, changes in skin color and texture, drooping eyelids, and changes in the shape of the lips and chin. Visual cues can be masked by cosmetics, medical procedures, lifestyle, and may also vary depending on ethnicity. This makes age determination from an image particularly challenging. Therefore, modern deep learning methods are appropriate for solving the age determination problem. This paper uses an approach based on convolutional neural networks to solve this problem. The following stages were completed for its implementation: obtaining and preparing data for analysis, selecting libraries and architectures for building convolutional neural network models, training convolutional neural network models, and evaluating the model performance. The following publicly available datasets were selected for the implementation of the first stage: Facial Age and UTKFace. A total of three data preparation scenarios were implemented: classification by exact age, classification by eleven age ranges, and classification by five coarse-grained age categories. The total number of images for classification was approximately 14,000, of which the training set contained 9,734 and the validation set contained 4,238 images. The Python programming language, as well as the TensorFlow, NumPy, Matplotlib, Seaborn, Scikit-learn, Pillow, and Tkinter libraries and frameworks, were used to build the neural network models. Two convolutional neural network architectures are considered and compared: ResNet50V2 and YOLOv8n. The YOLOv8n architecture demonstrated superiority over ResNet50V2 for each of the metrics: Accuracy, Macro F1-score, and Weighted F1-score. Issues with classifying middle-aged groups were identified, indicating the need for additional balancing of the training set and the use of augmentation methods.
Keywords:
INTELLIGENT SYSTEM, HUMAN AGE, NEURAL NETWORK MODELING, CONVOLUTIONAL NEURAL NETWORK, YOLOV8N, RESNET50V2, IMAGE CLASSIFICATION
INTELLIGENT SYSTEM, HUMAN AGE, NEURAL NETWORK MODELING, CONVOLUTIONAL NEURAL NETWORK, YOLOV8N, RESNET50V2, IMAGE CLASSIFICATION
