IMAGE RECOGNITION WITH DEEP NEURAL NETWORKS FOR PREDICTION OF RENTING RESIDENTIAL OBJECTS

Abstract

In this paper we research image recognition with deep neural networks for prediction of renting residential objects. The aim of this research is experimental evaluation of machine learning algorithims, with comparation of model results on initial data set and on extended data set, which is enriched with new attributes that are the result of object recognition on images of residential buildings. For the needs of research work, a project in Python, C# and SQL programming languages was made, which deals with predicting the interest class of tenants of housing in the city of New York in the United States on data from 2016., using machine learning algorithms. Possible tenant interest classes are: low, medium and high. Using pre-trained models of deep neural convolutional networks for image recognition, the initial data set is extended, with newly recognized objects as new attributes. After preparing the data for predictive modeling, the output class is predicted using machine learning algorithms for classification. In order to improve the accuracy of model prediction, models were evaluated experimentally and created by techniques for: fine tuning algorithm parameters, balancing data according to the output class and with technique for finding best K attributes. The results of the prediction on the initial and extended data sets are then compared. In conclusion, results and proposals for future application and development of machine learning algorithms are proposed.