本論文嘗試利用深度學習神經網路來提高靜態手勢的辨識率。整個系統架構分為影像前處理與神經網路訓練兩大部分。首先，將Moeslund的手勢資料庫提供之灰階影像處理過後，定位手勢位置並切割影像，再將其製作成訓練樣本與測試樣本。訓練的神經網路為堆疊降噪自動編碼器(Stacked Denoising AutoEncoder, SDAE)，是一種深度學習的網路架構，能夠學習到較好的特徵。訓練完成的SDAE可以進行辨識測試樣本的手勢為何種字母。在訓練時，我們還使用了Momentum下降梯度算法，能夠突破區域最小值並且加速收斂，以及Dropout能夠隨機選擇神經元讓其停止運作，降低過擬和(Overfitting)的發生機率。實驗使用1,440張影像訓練、600張影像測試，最後的辨識結果最高能達到99.96%的正確率。本論文提出的系統架構能夠快速且準確的辨識靜態手勢，具有高度的實用價值。 This study tried to utilize deep-learning neural networks to promote the recognition rate of static hand gestures. The recognition system is divided into two parts: image preprocessing and network training. First, the gray-scale images provided by Moeslund’s static hand gesture database are processed to locate the hand gesture subimages. Then, these gesture images are partitioned into two categories: training dataset and test dataset. The training neural network is Stacked Denoising AutoEncoders (SDAE), which is a deep learning network architecture that can learn better features. During training, the momentum descent gradient algorithm is used to break the region minimum and accelerate convergence. In addition, the dropout technique is used to randomly select neurons to stop functioning and reduce the probability of overfitting. In the experiment, 1,440 gesture images are used to train the network and 600 gesture images are used to test the network. The final recognition rate can achieve 99.96%. The recognition system proposed by this study can quickly and accurately identify static gestures, which is highly valuable and practical.
