Binarizing the 1938 architectural drawings of the Agriculture and Commerce Building with shallow convolutional autoencoders

Abstract

We binarized the historical 1938 architectural drawings of the Department of Agriculture and Commerce Building in Manila, which had become severely degraded over time. Because of the deep-seated stains on the original linen drawings, simple image processing techniques were ineffective in binarizing it. To address this challenge, we proposed using a shallow convolutional autoencoder to train a model that could automatically binarize the dirty drawings. Firstly, we manually binarized a portion of a drawing and used it as a training dataset for different CNN models. We achieved the best F1- and Intersection-over-union (IOU) scores of 0.98 and 0.95, respectively. Our results demonstrate that the autoencoder significantly outperformed the traditional thresholding technique. Additionally, we found that a shallow model is already sufficient for the task at hand.