Real estate property comparison in the greek market using advanced image similarity methods and web scraping techniques

Abstract

The dynamic and complex nature of the real estate market, especially in regions like Greece with its diverse platforms and non-standardized content, poses significant challenges in data collection and analysis. This thesis presents a comprehensive system that integrates advanced web scraping techniques, machine learning models, and a full-stack Django-based application to significantly enhance the collection, processing, and analysis of real estate data. Central to this system is an innovative image similarity model, designed to improve the detection and comparison of real estate properties based on visual content, thereby enabling a more sophisticated analysis of market dynamics. At the core of this system is the development of an image similarity model utilizing the ResNet50 architecture, optimized for visual recognition tasks within the real estate domain. The dataset, which includes images collected from Greek real estate platforms, is processed through a pre-trained ResNet50 model, fine-tuned to extract feature embeddings rather than perform direct classification. These images undergo preprocessing, including normalization and resizing to 224x224 pixels, to align with the input requirements of the ResNet50 model. The model then generates a 2048-dimensional feature vector for each image, effectively capturing its visual characteristics. These vectors are stored systematically for efficient retrieval and comparison in image similarity tasks. The system is fortified with robust data management techniques, such as checkpointing and error handling, ensuring reliable processing of large-scale datasets. By leveraging the pre-trained ResNet50 model, the system achieves high accuracy in image similarity tasks while minimizing computational overhead, offering a scalable and efficient solution for real estate image analysis.