?
Разработка архитектуры классификатора для оценки состояния объектов инфраструктуры с применением нейронных сетей
In recent years, with the advancement of deep learning and neural network methods, their application in geospatial analysis tasks has become particularly relevant. A key challenge in this field is assessing the state of urban infrastructure, including the classification of buildings by their functional purpose (residential, commercial, governmental, industrial). The use of neural networks significantly enhances the speed and accuracy of the analysis, identifying problem areas at the scale of individual streets and districts, as well as entire cities, thereby facilitating informed decision-making for urban improvement.
This work proposes an architecture that combines the collection of panoramic images from online maps with their classification to analyze districts based on their development type. The key stages of the research have been carried out, including data collection and annotation, the selection of pre-trained models, their adaptation to the classification task, and the evaluation of the classifier's accuracy. A comparative study of pre-trained convolutional neural network models, such as ResNet-50v2 and VGG19, was conducted to build the classifier. Training was performed using the backpropagation method on a custom dataset created from open sources of city panoramas.
Additionally, the potential for integrating Natural Language Processing (NLP) and Vision-Language Model (VLM) methods to improve classification accuracy was investigated. Among the considered VLM models, particular attention was given to PaliGemma 2, which demonstrates high effectiveness in multidisciplinary analysis tasks. The classifier's results are presented using city districts as an example, confirming its practical applicability and high assessment accuracy. This work contributes to the further development of automated urban infrastructure analysis methods and highlights the promise of using modern neural architectures in this field.
In book
