| Project/Area Number |
23K16903
|
|---|
| Research Category |
Grant-in-Aid for Early-Career Scientists
|
| Allocation Type | Multi-year Fund |
|---|
| Review Section |
Basic Section 61010:Perceptual information processing-related
|
|---|
| Research Institution | Saga University |
|---|
Principal Investigator |
|
|---|
| Project Period (FY) |
2023-04-01 – 2026-03-31
|
| Project Status |
Granted (Fiscal Year 2023)
|
| Budget Amount *help |
¥4,550,000 (Direct Cost: ¥3,500,000、Indirect Cost: ¥1,050,000)
Fiscal Year 2025: ¥1,430,000 (Direct Cost: ¥1,100,000、Indirect Cost: ¥330,000)
Fiscal Year 2024: ¥1,430,000 (Direct Cost: ¥1,100,000、Indirect Cost: ¥330,000)
Fiscal Year 2023: ¥1,690,000 (Direct Cost: ¥1,300,000、Indirect Cost: ¥390,000)
|
|---|
| Keywords | Historical buildings / Machine learning / Image recognition / CNN / 都市景観評価 / クラスタリング / 編年 / 逆強化学習 / 建築史学 |
|---|
| Outline of Research at the Start |
この研究は、機械学習を用いた景観デザインにおける建物外観の評価支援学習システムの開発を目的としている。建築史学の手法と機械学習を組み合わせ、建物外観の設計に逆強化学習を適用するための基盤を構築することを目指している。未指導学習によるクラスタリングの結果と建築専門家による年代順の分類との差異を定性的に分析し、逆強化学習の報酬基準を決定する。現在は景観審議会の委員によるあいまいな基準で判断されている景観デザインプロセスを革新し、恣意性を排除し、建物分類にかかる時間を短縮し、人間には識別できない新しい分類基準を開発することを目的としている。
|
| Outline of Annual Research Achievements |
The research aimed to enhance image recognition of historical buildings using machine learning, focusing on districts with limited building counts. It addressed challenges in bias removal, exposure correction, and attention acquisition. By preprocessing images and applying Convolutional Neural Networks (CNNs) like ResNet-50, accuracy improved. Utilizing Grad-CAM quantified AI's focus on building parts, highlighting the impact of preprocessing. The study demonstrated potential in acquiring attention through AMs for human intervention, despite needing accuracy improvements. Future challenges include refining attention acquisition and expanding its application.
|
| Current Status of Research Progress |
Current Status of Research Progress
3: Progress in research has been slightly delayed.
Reason
The research progress has experienced a slight delay. Initially, the methodology hinged on employing Heatmap on IIC (Invariant Information Clustering) for attention acquisition. However, we encountered limitations. Specifically, the application of Heatmap on IIC proved unfeasible due to technical constraints. Consequently, we were compelled to explore alternative methodologies to achieve our objectives effectively. We transitioned to utilizing ABN (Attention Branch Network) as our primary method for attention acquisition. This decision was informed by the versatility and robustness of ABN in capturing attentional mechanisms within complex datasets.
|
| Strategy for Future Research Activity |
In our future work, we aim to address the challenges encountered and further enhance the efficacy of our research methodology. Firstly, we plan to refine and optimize the ABN framework for better attention acquisition, ensuring more accurate and comprehensive analysis of historical building images. Additionally, we intend to explore the integration of advanced image processing techniques to improve the preprocessing stage, mitigating biases and enhancing the quality of input data. Moreover, we aspire to collaborate with domain experts to validate our findings and refine our model's interpretability, ensuring its relevance and applicability in real-world contexts. Lastly, we envision disseminating our research outcomes through academic publications and conferences.
|
