以深度學習分析摺紙照片推估3-6歲兒童年齡

Using Deep Learning to Analyze Pictures of Origami to Predict the Age of Three-to-Six-Year-Old Children

游硯婷(Yen-Ting Yu) ； 林恭宏(Gong-Hong Lin) ； 黃千瑀(Chien-Yu Huang) ； 謝清麟(Ching-Lin Hsieh)

《職能治療學會雜誌》 39卷2期 (2021/12) Pp. 211-231

https://doi.org/10.6594/JTOTA.202112_39(2).0005

摘要

職能治療師常依兒童之活動作品品質判斷其發展狀況。然而，此判斷多為主觀質性描述，缺乏客觀量性數據之支持而易有判斷偏差。人工智慧（artificial intelligence, AI）可應用於分析兒童活動作品照片，透過電腦視覺擷取照片特徵，將這些特徵進一步進行特定變項（例如兒童年紀或發展表現）的預測／推估，則活動作品便能透過此預測／推估獲得一客觀的變項數值。因此，本研究以摺紙為例，以AI深度學習模型分析摺紙作品照片，使其推估兒童之生理年齡，以及找出較佳之照片拍攝角度。本研究共蒐集119名3-6歲兒童之摺紙作品。每件摺紙作品拍攝八張不同角度之照片。每種角度之照片皆分別進行AI深度學習模型訓練與驗證二階段。於訓練階段，我們隨機選取95筆（80%）資料，以深度學習模型（ResNet50）提取摺紙照片特徵，再將提取特徵以全連結層推估兒童生理年齡。於驗證階段，我們將訓練完成之模型套用至剩餘24筆（20%）資料。我們以決定係數（R^2）作為驗證模型推估力之指標。研究結果顯示，正面0度之照片拍攝角度具較佳的推估能力（驗證階段R^2=0.69）。結果支持摺紙作品可透過AI深度學習模型推估兒童之生理年齡。未來可增加樣本數以提升模型推估能力，並進一步利用摺紙作品於推估兒童發展程度（如動作發展）。

關鍵字

深度學習 ； 學齡前兒童 ； 兒童發展 ； 摺紙

並列摘要

Occupational therapists often evaluate children's development by assessing the quality of their works. However, these evaluations tend to be subjective and qualitative and lack the objective evidence-based support. Artificial intelligence (AI) models can objectively extract features from photos of children's work and then predict/estimate the specific variables (e.g., children's age or standardized scores of developmental tests). Therefore, this study aimed to apply AI to estimate children's age and find the angle of photos with the best estimation. A total of 119 children aged 3- to-6 years were recruited in this study. Eight photos of different angles of their origami works were collected. In the training phase, we randomly selected 95 sets (80%) of data to extract the features of photos with the ResNet50 model, and then to estimate children's age with the fully connected layer. In the validation phase, we analyzed the remaining data (24 sets, 20%) to estimate the power of these models. Higher coefficient of determination (R^2) indicated greater power of estimation. The photo of 0 degrees front side showed the best estimation power (R^2=0.69). Our results support that children's origami works can be used to estimate their age with AI deep learning models. In future studies, researchers should increase the sample size to improve the estimation power of these models and expand the application of origami works to estimate children's development (e.g., motor development).

並列關鍵字

Child Development ； Deep Learning ； Origami ； Preschool Children