机器学习视野下的建筑风格形式特征研究方法初探

Preliminary Study on the Research Methods of Architectural Style Form Features from the Perspective of Machine Learning

对不同时期的不同建筑风格形式特征的研究,一直是建筑历史与理论研究中的一个重要领域,长久以来一般采用"案例+分析+归纳"的方式进行。虽然通过这种方法已经基本确立了中外建筑史的多个典型风格时期,但由于历史的连续性与地域文化的交融性,在对实际历史建筑进行保护与改造时,往往很难给出明确的风格定位,或者对于同一个建筑上叠加的、残留的形式特征难以做出准确的判断,不利于后续设计工作的展开。当今人工智能领域中的机器学习方法,为改善上述问题提供了新的视角。本研究从该方法入手,探索如何通过不断提升机器对风格图片的分类能力,实现定量化的风格判断,适用于各种混合风格的情况,帮助设计者锁定各种形式风格的特征,为后续设计改造提供参考。这里以对巴洛克、拜占庭、哥特3种建筑形式风格的机器学习为例,介绍了如何通过学习3331张网络爬取样本,达到80.8%的分类正确率。同时,通过试验还对比了人工判断的71.0%平均正确率,以及利用迁移学习的特点在其他非建筑领域取得的81.8%正确分类效果。上述案例虽然能力有限,但已展示出机器学习方法对于建筑风格研究的独特价值,其方法具有普遍意义。

The study on the formal characteristics of different architectural styles in different periods has always been an important field in the study of architectural history and theory.For a long time,the method of"case+analysis+induction"has been generally adopted.Although,in this way,several typical style periods in the history of Chinese and foreign architecture has been basically established.But as a result of the interaction of the continuity of history and regional culture,in the actual protection and renovation of historical building,it is often difficult to make an accurate judgement on the style of the overlaid or residual form characteristics of the same building,which is not conducive to the development of subsequent design work.Machine learning methods in today’s artificial intelligence field provide a new perspective for improving the above problems.Starting from this method,this study explores how to continuously improve the machine’s ability to classify style images to achieve quantitative automatic form style judgment,which is applicable to various mixed styles.It helps the designer to lock the local features of various forms and styles and provides a reference for subsequent design and transformation.Taking machine learning of three architectural styles of Baroque,Byzantine and Gothic as an example,this paper introduces how to achieve 80%classification accuracy by learning 3331 network crawling samples.At the same time,the average accuracy rate of human designers of 71.0%and the 81.8%accuracy rate classification used the model in other fields are also compared through experiments.The above cases,though limited in capability,have demonstrated the unique value of machine learning methods for architectural style research,and their methods are of universal significance.