.png){kind=icon}
.png){kind=icon}
Return
Ancient architecture image classification with progressive stacking pseudoinverse learning
.png){kind=icon}
OA
AI
DOI：10.1038/s41598-026-44876-9.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In the task of ancient architecture image classification, due to the unconstrained random initialization of the weight matrix, the existing random projection branches cannot effectively focus on the key feature dimensions such as roof contour and bucket arch structure. This not only weakens the core basis for classification but also impairs the ability to distinguish between similar building variants. Meanwhile, the training mode of inputting all samples at one time is easy to cause the model to over-fit the architectural style of specific regions and weaken the cross-regional generalization performance. To address these issues, we propose an ancient architecture image classification with progressive stacking pseudoinverse learning (AAPSP). It comprises two core modules: the key features stacking pseudoinverse learning of ancient architecture (KFSP) and the progressive optimization learning model of ancient architecture samples (POL). Specifically, KFSP initializes the weight matrix to a specific pattern that conforms to the distribution of architectural features (such as Gaussian distribution focusing on the roof contour, uniform distribution capturing decorative patterns). By combining the attention mechanism (AM), a higher weight is assigned to the base learner of key component recognition to optimize the feature space modeling. POL employs a dynamic sample screening strategy to preferentially select sample iterative optimization models with rare features to reduce redundancy and enhance generalization ability. Experiments on six types of Chinese ancient architecture datasets show that AAPSP performs well in accuracy, precision, recall rate and F1 score.
Keywords:
Ancient architecture image classification
Stacking pseudoinverse learning
Key features
Sample progressive optimization
Journal
IF:
3.9
Papers:
27.8W
Citations:
83.5W