Real Estate Image Classification For E-commerce Website

Abstract

Currently, the image is the most important part of e-commerce websites and has a very significant impact on the real estate industry. In this work, we represent a novel approach for image classification in real estate. This problem can be applied to resolve some automated functions in operating the website when using machine learning algorithms in data-driven approaches. We can find some datasets about real estate images, such as REI and SUN, for this classification problem. However, they do not have images related to ecommerce websites such as TopReal in the Vietnam market. Therefore, we propose a new dataset called REID that is based on real requirements from the TopReal website and the others. This dataset has a hierarchy with 4 generics and 8 specific classes. To this end, we propose two model-based ConvNets for classification problems. In the first model, we design based on the LeNet architecture as in the baseline model to investigate and analyze the dataset for deep understanding. In the second model, we propose a transfer learning-based approach with ResNet as the feature extraction module and SVM as the classifier. The experimental results show that we obtain 97.6% and 87.3% accuracy for classes 4 and 8, respectively, on the REID dataset. These results show that our proposed methods achieve high accuracy on a real estate image dataset that adapts the websites in the Vietnam market and can extend easily to the others. Furthermore, the REID dataset can be extended to resolve many problems in real estate images for the Ecommerce Website.