ArchDetect: Automated Architectural Analysis

Updated 21 January 2026

ArchDetect is a class of automated systems that extract architectural and structural entities using data-driven methods across domains like art history, archaeology, and software engineering.
It employs core methodologies such as deep convolutional neural networks, NLP-based extraction, and 3D point cloud analysis to transform raw data into structured representations.
Recent implementations demonstrate high accuracy and modular design, enabling efficient applications in archaeological documentation, architectural style clustering, and software traceability.

ArchDetect refers to a class of automated or semi-automated systems designed for architectural analysis, detection, and understanding in domains such as visual art history, archaeology, software engineering, and computer vision. ArchDetect systems share the goal of extracting explicit architectural or structural entities (e.g., building typologies, archaeological objects, architectural design features, software components, or defect features) from large and heterogeneous data modalities, using data-driven, statistical, or deep learning techniques. Implementations span computer vision, natural language processing, geometric/statistical modeling, and hybrid, multimodal pipelines.

1. Core Methodologies Across Domains

ArchDetect methodologies vary by application area but are unified by their focus on transforming raw sensory or textual data into structured representations of architectural entities.

Computer Vision for the Built Environment and Archaeology: Deep convolutional neural networks (DCNNs), notably NASNet backbones, are trained for visual classification of building designs by architect, achieving measurable similarity scores and quantitative clustering of architectural styles. Object detection methods, such as Faster R-CNN with MobileNetV3 backbones, are employed for archaeological entity extraction from scanned documentation, followed by post-processing (e.g., non-maximum suppression, association of artefacts to graves) and downstream spatial/geometric analysis (Yoshimura et al., 2018, Klein et al., 2023).
Textual/NLP-based Architectural Entity Extraction: In software engineering, LLMs drive tools like ExArch (for simple Software Architecture Model generation) and ArTEMiS (for Named Entity Recognition and traceability between documentation and code), combining prompt engineering, similarity-based matching (Jaro-Winkler, Levenshtein, text-embedding distance), and multi-source aggregation. These systems algorithmically recover architectural entities from documentation and/or code, automating traceability link recovery with state-of-the-art F₁ performance (Fuchß et al., 4 Nov 2025).
3D Point Cloud Feature Analysis: In infrastructure anomaly detection, customized multimodal feature descriptors (e.g., 3DMulti-FPFHI) fuse geometric (Fast Point Feature Histogram—FPFH) and intensity information. These features integrate with anomaly detection frameworks (PatchCore) to enhance the localization of fine cracks or material anomalies in 3D mesh data (Jing et al., 9 Feb 2025).
Shape and Contour Analysis in Archaeology: Techniques include contour detection (OpenCV findContours), geometric attribute extraction (measuring grave orientation, area via minimum-area rectangles), and dimensionality reduction for morphometric analysis (Elliptic Fourier Descriptors, PCA) (Klein et al., 2023).

2. Notable Architectures and Algorithms

Several architectures and specific algorithmic design choices typify current approaches:

DCNN-Based Style and Author Classification: NASNet-Large backbones, pretrained on ImageNet and fine-tuned on custom labeled datasets, leveraging global average pooling and cosine distance for embedding-based visual similarity and clustering architectures (k-means in PCA-projected space). Training regimes emphasize data augmentation and regularization for generalization (Yoshimura et al., 2018).
Object Detection in Archaeological Documentation: A Faster R-CNN architecture, with a MobileNetV3-Large backbone, trained for 13 object classes, achieves high precision and recall on legacy PDF-derived archaeological datasets. Supporting modules include ResNet-152 classifiers for north-arrow orientation and skeleton pose type; scale calibration is achieved through OCR (Tesseract) and pixel-to-real conversions (Klein et al., 2023).
LLM-Driven SAD and Code Analysis: ExArch leverages deterministic temperature=0 LLM prompts and aggregation, using normalized Levenshtein similarity to reconcile component names. ArTEMiS applies chained prompts (task description and output formatting), together with multi-metric entity matching, for flexible SAD→SAM trace link generation, closely matching manual state-of-the-art at substantially reduced human labor (Fuchß et al., 4 Nov 2025).
Edge Detection in AR for Architectural Features: Novel convolutional kernels (weighted for 0°/90° edges) augment building detection performance. Kernel design tunes trade-offs between detail preservation and noise suppression; performance is quantified via edge pixel/segment statistics (Orhei et al., 2021).

3. Empirical Performance and Evaluation

ArchDetect systems have demonstrated state-of-the-art or near-manual level accuracy in multiple benchmarks:

System/Domain	Task/Metric	Reported Value(s)
DCNN (NASNet)	Architect classification (top-1)	73.17% (Yoshimura et al., 2018)
DCNN (NASNet)	Top-5 accuracy	87.07% (Yoshimura et al., 2018)
ExArch + TransArC	SAD→code TLR, weighted F₁	0.86 (manual SAM: 0.87) (Fuchß et al., 4 Nov 2025)
AutArch (Faster R-CNN)	Grave detection, precision/recall/F₁	100%/100%/1.0 on selected test sets (Klein et al., 2023)
DArch (Dental arch)	Tooth centroid detection accuracy	99.7% (APS) vs. 84.3% (FPS) (Qiu et al., 2022)
DArch	Segmentation Dice (weakly supervised)	97.4% (Qiu et al., 2022)

Evaluation protocols typically emphasize cross-domain generalizability (e.g., tested on out-of-sample publications in AutArch), per-class precision/recall for cluster fidelity in DCNN approaches, and micro/macro-F₁ for traceability in code analysis tools.

4. Practical Integration and System Design

ArchDetect system integration follows modular and reproducible patterns:

Pipeline Modularity: Backends commonly expose core detection/classification services (e.g., via REST APIs), decoupling pre-processing, detection, feature extraction, and user-in-the-loop validation (AutArch web GUI, manual contour adjustments) (Klein et al., 2023).
Data Ingestion: Interfaces ingest diverse source formats: raw images (building classification), legacy PDFs (archaeological object detection), structured documentation (software engineering), and 3D point clouds (infrastructure).
Manual Validation: GUI workflows enable domain experts to validate/correct automated predictions, which can then be used for incremental fine-tuning of models in a human-in-the-loop paradigm (Klein et al., 2023).
Production Recommendations: Guidance includes prompt and threshold tuning in text-based systems, caching and incremental regeneration of models upon data changes, and fallback to classical approaches where LLMs or deep learning are less robust (Fuchß et al., 4 Nov 2025).

5. Comparative Analysis Across Application Areas

ArchDetect serves as an umbrella term for several parallel, domain-specific innovations:

Art History and Visual Design: Image-based clustering of architectural works, recovering art-historical groupings while uncovering new, machine-learned clusters based on DCNN-learned visual features (Yoshimura et al., 2018).
Archaeological Data Standardization: Automated extraction of structured, spatially annotated object records from heterogeneous legacy sources, offering significant speed and error-reduction over manual approaches (Klein et al., 2023).
Software Architecture and Traceability: LLM-driven automation of component recognition and link recovery, closing gaps between unstructured documentation and maintainable, up-to-date architecture models (Fuchß et al., 4 Nov 2025).
3D and Defect Detection: Multimodal features and integrated anomaly detection for infrastructure health monitoring, with code and data available for reproducibility (Jing et al., 9 Feb 2025).

A plausible implication is that core methodological advances—such as embedding-based similarity, multimodal data fusion, and human-in-the-loop review—are highly transferrable between these settings, even though object classes and data modalities differ fundamentally.

6. Strengths, Limitations, and Future Prospects

Strengths:

Minimal human annotation for high-level accuracy, especially with LLM-assisted extraction and weak supervision (Fuchß et al., 4 Nov 2025, Qiu et al., 2022).
Modularity and reusability of core detection, validation, and association components enable rapid adaptation to new object classes or research domains (Klein et al., 2023).
Quantitative performance is competitive or superior to manual and heuristic approaches across multiple domains.

Limitations and Open Challenges:

Domain-specific training data remains essential for rare or complex object classes (e.g., new archaeological artefacts, specialized software patterns).
Output quality is sensitive to input quality (e.g., degraded scans in OCR, missing or informal documentation in SAD).
Compute and latency costs may rise substantially with larger foundation models or in high-resolution imagery.

Future Directions:

Combining vision and language modalities (e.g., integrating tabular data parsing with image analysis) is highlighted as a path for richer and more automatic knowledge extraction (Klein et al., 2023).
Advances in transformer-based object detection, more robust OCR, and improved prompt/dataset synthesis promise further gains in both accuracy and applicability.

7. Representative Implementations and Open Resources

Several publicly available implementations and datasets are aligned with ArchDetect research:

System	Domain	Resource URL
3DMulti-FPFHI	Infrastructure/3D	https://github.com/Jingyixiong/3D-Multi-FPFHI (Jing et al., 9 Feb 2025)
AutArch	Archaeology	(As described in (Klein et al., 2023); contact authors for code)

For DCNN-based architectural style clustering, protocols, code, and pretrained weights follow the procedure outlined in "Deep Learning Architect" (Yoshimura et al., 2018). AutArch’s pipeline and web GUI design documented in (Klein et al., 2023) can be extended for new annotation and detection scenarios, illustrating the generalizability of the modular ArchDetect paradigm.

References:

(Yoshimura et al., 2018): Deep Learning Architect: Classification for Architectural Design through the Eye of Artificial Intelligence (Orhei et al., 2021): A Novel Edge Detection Operator for Identifying Buildings in Augmented Reality Applications (Qiu et al., 2022): DArch: Dental Arch Prior-assisted 3D Tooth Instance Segmentation (Klein et al., 2023): AutArch: An AI-assisted workflow for object detection and automated recording in archaeological catalogues (Jing et al., 9 Feb 2025): A 3D Multimodal Feature for Infrastructure Anomaly Detection (Fuchß et al., 4 Nov 2025): Who's Who? LLM-assisted Software Traceability with Architecture Entity Recognition