Deep Learning-Based Building Image Information Extraction and Design Similarity Pre-Review Framework for Sustainable Architectural Design Decision Support

Design similarity disputes in architecture can lead to injunctions, damages, redesign, and, in extreme cases, the demolition and reconstruction of completed buildings, wasting resources, cost, time, and embodied carbon. Reviewing design similarity at the early design stage is therefore a prerequisite for sustainable design decision-making, yet conventional review relies on qualitative expert judgment with limited objectivity and repeatability. This study proposes a deep learning-based building image information extraction and design similarity pre-review framework that extracts building information from images and retrieves visually similar existing buildings as reference cases, rather than determining legal plagiarism. The implemented framework consists of a data module, a building module, and a prototype review interface; exterior material, structural, and frame modules are defined as conceptual extensions. The data module refines building images collected by web crawling and direct photography using rule-based filtering, DINOv2 feature extraction, and K-Means clustering. The building module combines SAM, ResNet-50, and DeepLabV3+ to identify building objects and extract building regions and contours. From 42,172 collected images, 10,026 were retained. ResNet-50 reached a peak Top-1 accuracy of approximately 92%, and DeepLabV3+ achieved approximately 89% aACC and mACC and 81% mIoU. The results demonstrate the feasibility of image-based similarity pre-review as a digital foundation for reducing design rework and supporting sustainable design decisions.