Pilot experiment DSS Hasmo/Herod fraglet-based dataset bipartitioning (70x70 Kohonen map)

As a follow up on a pilot experiment for manuscript dating in the DSS/IAA ERC project, 590 scans, each labeled on global style (Hasmonean/Herodian) were binarized using Maruf Dhali's BiNet neural network. Fragmented connected component contours were computed and compared to a precomputed <strong>70x70</strong> Kohonen map of DSS fragmented connected-component contours (*.fco3). Kohonen cells obtained the frequency counts for the labels Hasmonean and Herodian. Individual scans were characterized by their total Hasmo, Herod counts. Augmentation using random elastic morphing was used to obtain a comparable number of FCO3s per scan. The 2D feature vector with total counts #Hasmo and #Herod was subjected to PCA, with the goal of identifying the most informative axis for the Hasmonean/Herodian distinction. 91% of the samples (scans) could be correctly classified. Although this is a simple, within dataset bipartitioning experiment, the good results were taken as an indicator that the FCO3 fraglet feature would be useful in a more fine-grained style-based manuscript dating attempt using the radiocarbon-labeled data. See https://zenodo.org/deposit/8380279 for an accompanying .pdf technical report.作为DSS/IAA ERC项目中古手稿断代预实验的后续研究，590份基于整体风格（哈斯蒙尼王朝（Hasmonean）/希律王朝（Herodian））标注的扫描件，通过Maruf Dhali提出的BiNet神经网络（BiNet）完成二值化处理。研究人员计算了每份扫描件的碎片化连通分量轮廓，并将其与预先构建的70×70尺寸DSS碎片化连通分量轮廓科赫农映射表（Kohonen map，*.fco3）进行比对。每个科赫农节点（Kohonen cells）均统计了哈斯蒙尼王朝与希律王朝两类标签的出现频次，每份扫描件的特征由其对应两类标签的总频次共同表征。研究采用随机弹性形变的数据增强手段，使每份扫描件对应的FCO3样本数量趋于一致。将包含#Hasmo与#Herod总频次的二维特征向量代入主成分分析（PCA），旨在找出区分哈斯蒙尼王朝与希律王朝风格最具信息量的坐标轴。最终91%的样本（扫描件）实现了正确分类。尽管这仅是一项基于数据集内部的简单二分划分实验，但优异的分类结果表明，FCO3微片段特征可在后续基于放射性碳标定数据的细粒度风格化手稿断代研究中发挥应用价值。相关配套PDF技术报告可通过链接https://zenodo.org/deposit/8380279获取。