AIC's 51st Annual Meeting

The National Museum of Taiwan Literature (NMTL) is Taiwan’s first national-level literary museum. The majority of the collections in NMTL are paper-based materials such as manuscripts and books. When the collections exceeded 110,000 items, its condition assessment and documentation work becomes more cumbersome and time-consuming. In order to relieve the current situation, the NMTL launched an artificial intelligence (AI) learning research project in cooperation with the Industrial Technology Research Institute (ITRI) in 2021, for the identification of deterioration on paper. Extending from the main axis of museum collection and preservation to the application of new technology, the first stage of this research was to collect representative images of deterioration from manuscript and book collections. These images of deterioration were used to train the AI model through a co-creation competition activity, and the feasibility of AI identification of deterioration has been preliminarily verified.

However, several difficulties were also faced during the AI recognition process. Image data is complicated due to the fact that the diversity of paper substrate background interferes with the deterioration target, which increases the difficulty for accurate recognition. Even the same type of deterioration may have further classifications, and different types of deterioration show very different visual patterns, resulting in a highly complex image that restricts the performance of AI learning. Sufficient quality data collection and input is highly influential towards AI learning quality.

To resolve this, a further research was launched to find the methodology to construct a proper image database for AI learning. Firstly, image characteristic analysis such as contrast and brightness level were conducted to extract the effectiveness of training data from deterioration images. Secondly, the visual patterns of deterioration were then divided into several classes to help improve AI’s learning quality. The analysis method was implemented as operation criterion to improve the quality of labeling data. Finally, the image data collected from the analysis results went through a data augmentation procedure, and were verified with experimental training and testing for AI recognition performance. The system parameters, photography equipment and labeling software were then fine-tuned to meet the standard of NMTL’s digitalization workflow. Reason being, the learning quality of AI is highly dependent on sufficient quality data input.

In this research, automated visual recognition technology for deterioration on manuscripts was verified, and the methods for collecting AI learning data are also evaluated. It is hoped that the maturing AI visual analysis technology and machine learning model can assist collection condition assessment in the future, so as to enhance efficiency and record documentation more comprehensively compared to traditional manual procedures. Through the interdisciplinary research and fusion with AI technology, the digitization and documentation of collections will be implemented to help the NMTL further understand the overall condition of its collections, effectively integrating the accumulated data from the past and the continuous documentation output of the future.