人臉屬性偵測基於深度神經網路

Abstract

近年來，人臉屬性在電腦視覺領域越來越受到重視，包括辨識、分類、檢索等研究方面。在照片中偵測人臉屬性是富有挑戰性的，因為真實世界的背景雜亂或是人臉角度變化影響黑大，例如:大小、姿勢、光影變化。有效解決這個問題的辦法是建立適當的特徵去表示人臉。近幾年在影像分類領域崛起的Deep Neural Network (DNN)已被大量應用於此類問題。DNN可以有效的解決這個問題，但是DNN需要大量的計算量及空間需求量。 增進人臉屬性分類的準確率是很重要的一步。所以我們的目標是用比較小的DNN模型，降低在偵測時的運算量及所需要的儲存空間。我們的模型是從資料中去學習人臉屬性之間的關係，先從大量的人臉識別中學習，再用人臉屬性調變參數。利用參數數量較少的DNN模型來偵測人臉屬性仍可以達到與傳統方法近乎相同的準確率。我們的人臉屬性偵測的實驗結果表現於現有公開最大量的人臉屬性資料庫CelebA。

Facial attributes have gained popularity in the past few years in machine vision tasks including recognition, classification, and retrieval. Predicting facial attributes from web images is very challenging due to background clutters and face variations, such as scale, pose, and illumination in the real world. The key to this problem is to build proper feature representations to cope with these unfavourable conditions. Given the success of deep neural network (DNN) in image classification, the high level DNN feature as an intuitive and reasonable choice has been widely utilized for this problem. DNN is powerful to handle face variation, but it needs heavy computation efforts and memory storage resources. Improving the accuracy of attribute classifiers is an important first step in any application which uses these attributes. Therefore, our goal is improving face attribute detection performance with smaller architecture of deep models. Our network is pre-trained with massive face identities, then fine-tuned with attribute labels. We consider the DNN features as face representation for attribute prediction. We demonstrate the effectiveness of our method by producing results on the challenging publicly available datase CelebA.