FLEXNet: Learning Deep Features Using Autoencoders for Floor Plan Image Retrieval

Content-based retrieval of scanned document images is gaining research. A system that automatically analyses and retrieves similar floor plan images depending on user needs might help architects create new ideas and provide client recommendations. Our model, FLEXNet (Floor pLan EXplorer Network), extracts deep features from images using an Autoencoder-based approach and facilitates exploration of a database of floor plan images to retrieve similar images using a K-nearest neighbour classifier. We evaluated FLEXNet on images perturbed with rotation, scaling, and noise in a publicly available dataset. FLEXNet outperforms state-of-the-art methods with an 85% mean average accuracy.