基於深度學習於低照片品質下的行人特徵辨識

Abstract

行人特徵辨識在電腦視覺領域中一直是個很重要且對人類社會有價值的問題，因為其應用廣泛，從安全領域到商業領域都有其應用價值。而行人姿勢、照片光線、背景複雜、細微特徵的問題都使得行人特徵辨識這個問題的難度更大。目前已經有許多研究都提出相對應的解決方法來處理上述的問題，但都忽略了從低成本的監視器的獲取的相片的品質是遠低於一般相機的。而從其他研究中，我們可以得知相片品質是會影響機器無法習得穩健的特徵以進行正確的分類。在這篇研究中，我們透過增加機器學習的資訊量，並讓機器自己去選擇對自己學習有利的資訊，屏除不利於學習的部分，重新組合成最適合機器去學習的相片。在這樣的機制底下，我們可以減低照片品質的影響，例如雜訊，藉此讓機器可以習得更穩健的特徵，以達到更高的分類準確度。我們將我們提出的網路架構實驗在目前行人特徵辨識最大的兩個資料集上 (PA-100K, RAP)，透過一系列的實驗去證明我們提出的架構確實可以幫助提高機器分類的準確度，也可以有效地減低雜訊的影響並維持一定的分類準確度，在消融實驗中也可以佐證我們架構中的每個部份都有利於機器分類的準確度。從實驗中也可以觀察到，我們的方法用於一般的分類網路上即可勝過目前在行人特徵辨識問題中表現最好的方法，而我們的方法更可進一步地用於目前表現最好的分類網路上，達到更高的準確度。

Pedestrian attribute recognition is an important and valuable task in computer vision field attributed to its extensive application, such as person retrieval with attributes, marketing strategy building and person re-identification. However, it is also a challenging task due to various viewpoints, poses, illumination, backgrounds and fine-grained attributes. Although many methods have been proposed in order to deal with these issues, they neglect low image quality issue which often occurred in surveillance camera. Dodge also clarify in their work that image quality will affect machine do classification. To handle this issue, we propose a way to increase more samples and make model to learn how to select useful region in different images in order to combine a new image for more efficient learning. In this way, our model can reduce the influence of low image quality (e.g. noise) and learn the more robust features for more accurate classification. We evaluate on two biggest pedestrian attribute recognition datasets (PA-100K, RAP) through a series of experiments and ablation studies to verify our model can improve the classification accuracy further and showcase the effectiveness of the proposed architecture. Experimental results also demonstrate that our method which add on the common classification networks can outperforms other state-of-the-arts. Furthermore, our method can add on the state-of-the-arts and improve the accuracy further.