A Case Study on Architectural Sketch Recognition Utilizing Deep Learning Networks for Exterior and Interior Datasets

Abstract

Sketching is a pivotal component in facilitating the effective conveyance of ideas and the actualization of architectural design concepts. The potential applications of machine learning and computer vision algorithms in the fields of technical drawing and architectural graphic communication are substantial, presenting a diverse array of possibilities. This research investigates the effectiveness of deep learning-based classification techniques in analyzing both indoor and outdoor freehand architectural perspective drawings. Furthermore, the transfer learning approach was employed in this binary classification problem. The primary aim of this study is to train deep neural networks to recognize and interpret freehand architectural perspective drawings effectively and precisely. In this context, pre-trained models such as GoogLeNet, ResNet-50, AlexNet, ResNet-101, Places365-GoogLeNet, and DarkNet-53 were finetuned. The findings indicate that the ResNet-101 architecture has significant levels of validation accuracy, yet the validation accuracy of the Places365-GoogLeNet and AlexNet pretrained models is comparatively lower.

Keywords: Machine Learning, Transfer Learning, Drawing Recognition, Deep Neural Nets, Image Classification