以卷積神經網路及資料擴增方式訓練識別低應 力響應物件形貌人工智慧模型之方法

隨著人工智慧技術躍進，觸覺感測已成為智慧系統於複雜環境中進行互動與感知的關鍵技術。然而， 低成本平面陣列感測器普遍面臨空間解析度受限以及物體重量與幾何形狀特徵高度耦合等挑戰，導致辨 識精準度難以提升。本研究提出一種觸覺辨識運算框架，採用16 × 16 軟性薄膜壓力感測陣列，依序透過 數據收集、特徵解耦、擴增建模及智慧辨識之系統化流程進行分析，並透過壓力總和標準化消除絕對重 量干擾，強制模型優先提取幾何特徵。在模型評估方面，提案方法開發一套整合空間位移、幾何旋轉、 插值縮放與隨機高斯雜訊之十倍資料擴增流程。此外，本方法亦針對33 種常用機器學習演算法與最佳化 之輕量化卷積神經網路建立完整測試。實驗結果顯示，相較於傳統模型在一般測試中之差強人意表現， 所提出之輕量化卷積神經網路架構於資料擴增後，在新舊測試集中均達成100% 辨識準確率。本方法透過 數位訊號處理與深層表徵學習實證，可補償低解析度訊號之限制，為智慧觸覺辨識應用建立一套標準高 效建模流程。

Abstract：With artificial intelligence evolving at breakneck speed, tactile sensing has become a key for intelligent systems to interact with and perceive complex environments. However, cost-effective planar array sensors face challenges, such as limited spatial resolution and strong coupling between weight intensity and geometric features, they restrict the accuracy of object recognition. This study proposes a computational framework for tactile object recognition using a 16 × 16 flexible thin-film force sensor array for data acquisition. The system follows a systematic data acquisition, feature decoupling, augmentation, and recognition architecture. A forcesum normalization is implemented to eliminate absolute weight interference, forcing the model to prioritize the extraction of spatial geometric features. To address practical challenges, such as insufficient sample variations and sensor drift, a 10-fold data augmentation pipeline comprising spatial translation, geometric rotation, interpolative scaling, and Gaussian noise injection was developed to expand the training set. A comprehensive benchmark was established by assessing 33 traditional machine learning methods alongside a customized lightweight convolutional neural network (CNN) optimized for tensor-formatted data. Results show that while traditional models suffer significant accuracy deterioration when facing new test sets, the proposed CNN architecture consistently achieves 100% recognition accuracy across both original and new test sets after augmentation, which confirms that data augmentation and deep learning can effectively compensate for the hardware limitations of lowresolution sensors.