深度神經網路的訓練通常需要大量的訓練數據，但目前含有真實三維人臉模型的數據非常少，因此本論文使用的方法為弱監督式學習，結合卷積神經網路模型以及三維可變形模型針對圖像進行深度的特徵提取及回歸相關係數並利用可變形模型重建以及改善其前處理和損失函數，利用圖像高、低層次的資訊，在沒有真實標籤的情況下，能有效地重建圖像人臉，對於角度變化豐富或是含有遮擋之圖像也能獲得很好的重建品質，後續將在兩個數據集中評估及分析其實驗結果，並且展示重建之三維人臉模型。;Nowadays, due to the maturity of technology and the perfection of instruments and equipment, the application of many industries has gradually expanded from two-dimensional plane to three-dimensional space, such as video entertainment industry, face recognition, medical cosmetology and other related fields, in order to realize a more realistic environment and provide excellent sensory experience for the public.
Traditional face reconstruction algorithms are unable to learn real face features in depth and rely on low-level information of the image, while the software modeling method and instrument modeling method used by the industry are time-consuming and costly. Therefore, in recent years, 3D face reconstruction techniques based on deep learning have shown effective results in terms of quality and efficiency, balancing quality and time to reconstruct 3D face models corresponding to images.
Therefore, this paper uses a weakly-supervised approach that combines a Convolutional Neural Network and a 3D Morphable Face Model. The face can be effectively reconstructed without realistic labels by using the depth feature extraction and regress coefficients of the image, and improving the pre-processing and loss function. The reconstruction results will be evaluated and analyzed in two datasets, and the reconstruction results will be demonstrated.
