用於暗淡影像之深度人臉辨識

Abstract

當套用先進的深度人臉辨識模型於低光源下所拍攝到的影像時，會產生嚴重的人臉辨識準確率下降；主要是由於低光源下所拍攝的人臉影像特徵無法與正常光源下所拍攝到的人臉影像特徵被正確地匹配。在本研究中，我們首先評估一種較為直觀的方法，該方法在執行人臉辨識前會先增強人臉影像的對比度及亮度。然後，基於實驗結果，我們提出了一個新穎的深度人臉辨識架構來解決此問題.此架構由一個特徵恢復網路(feature restoration network)、特徵提取網路(feature extraction network)、特徵向量優化模組(embedding refinement module)，以及特徵向量匹配模組(embedding matching module)所組成。特徵恢復網路採用基於卷積神經網路的兩分支結構，從原始影像和亮度增強影像生成特徵恢復影像(feature-restored image)。 特徵提取網路將特徵恢復圖像以高維空間中的特徵向量來表示。該特徵向量由特 徵向量優化模組進行進一步優化，並由特徵向量匹配模組用於人臉驗證(verification)和識別(identification)任務中。我們提出的架構在 SoF 資料集上能將人臉驗證準確率提高 3.1%~9.1%。而在人臉識別任務中，可將 rank-1 的識別準確度提高 3.7%。

The performance of many state-of-the-art deep face recognition models deteriorates significantly for images captured under low illumination, mainly because the features of dim probe face images cannot match well with those of normal-illumination gallery images. In this thesis, we first evaluate an intuitive approach that enhances the illumination of face images before performing face recognition. Then, based on the results, we propose a novel deep face recognition framework to address this issue. The framework consists of a feature restoration network, a feature extraction network, an embedding refinement module, and an embedding matching module. The feature restoration network adopts a two-branch structure based on the convolutional neural network to generate a feature-restored image from the raw image and the illumination-enhanced image. The feature extraction network encodes the feature-restored image into an embedding, which is then made more discriminative by the embedding refinement module and used by the embedding matching module for face verification and identification. The overall verification accuracy is improved from 3.1% to 9.1% when tested on the Specs on Faces (SoF) dataset. For face identification, the rank-1 identification accuracy is improved by 3.7%.