Remoção de objetos em imagens do mercado imobiliário usando inpainting

• Autor(a) principal: Reis, André Ferreira Alves dos
• Data de Publicação: 2023
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10400.22/23868
• topic: Real Estate; Inpainting; Image Object Removal; Computer Vision; LaMa; Imobiliário; Remoção Objetos Imagem; Visão Computacional

The real estate market is one of the most valuable and influential markets in the world. An essential part of this market is the first contact with potential customers through property listings on real estate websites. To make listings more attractive, professionals often hire external services or use image editing software to improve the quality of the images shown. Real estate agents often use tools to remove unwanted objects from images to declutter rooms, draw attention to the property, and eliminate uncontrollable environmental elements. However, existing solutions are not cost-effective for real estate agencies, as they have to pay for each image and the results can take up to two days to be delivered. In recent years, the use of deep learning artificial intelligence algorithms has revolutionised image removal technology, with current technologies capable of producing natural and realistic results. This paper focusses on the development of a deep learning inpainting image object removal solution for real estate images to be integrated into Maxwork, the back-office portal used by RE/MAX Portugal, one of the largest real estate companies in Portugal. This solution can remove objects from real estate images and contains several additional features, including the ability to undo and redo, and compare the original image with the most recent result. It uses the LaMa inpainting deep learning model, which proved to be more effective than other state-of-the-art models. The effectiveness of the solution was evaluated with several objective metrics, such as Fréchet Inception Distance (FID) and Learned Perceptual Image Patch Similarity (LPIPS), and subjectively with user feedback that gave an average rating of 4.31 for ease of use and 3.90 for satisfaction with the results obtained. Furthermore, ethical considerations related to image editing in the real estate sector are discussed to ensure a transparent and honest use of the solution. This paper also outlines the details of an experiment to train a new deep learning inpainting model using a collection of around 100,000 real estate images from past sold listings. However, the results of the experiment showed that the trained model was not able to outperform the pre-trained LaMa model, as it scored worse across all metrics. Additionally, this paper provides research on existing real estate image object removal solutions, the current application fields of deep learning inpainting models, and the evolution of inpainting models from traditional methods to current state-of-the-art deep learning models.