低光照車輛再識別的自我增強學習

Abstract

在車輛重識別任務中，白天和夜晚的光照分布不同導致域差異，是使模型辨別車輛身分性能下降的一個重要因素。由於夜間圖像的能見度不足，增強低光照圖像是解決此問題的常見方法，然而這些方法存在著一些問題：低光照的汽車圖片缺乏正常光照圖像，無法適應於監督式學習的光照提升方法；而非監督式學習的光照增強方法容易使影像產生色偏，但車輛的顏色是重要特徵，色偏問題會使得模型辨認能力下降；此外，多數方法無法適應不同程度的光照，會大幅降低車輛重識別任務在實際上的應用價值。為解決這些問題，本文提出一種全新的自校正模塊，用於提高圖像的光照程度。該模塊可以直接安裝在重識別模型前面，與重識別模型聯合訓練，不需要正常光照圖像或額外的損失函數，即可實現圖像光照提升且不影響車輛原始顏色。該模塊可以同時適應白天和夜間的不同程度光照，不須考慮原始訓練資料的光照分布問題。此外，本文提出了一個全新的車輛重識別資料集，包含了白天和夜晚的訓練和測試資料集。通過實驗證明，我們所提出的方法能夠有效提高影像的光照度，並提高重識別網路在低光照情況下的表現能力。

In vehicle re-identification(ReID) tasks, the difference in illumination distribution between daytime and nighttime results in domain discrepancies, which is a significant factor leading to decreased model performance in identifying vehicle identities. Enhancing low-light images is a common method to solve this problem since nighttime images have poor visibility. However, these methods have some issues: low-light vehicle images lack normal lighting images, making them unsuitable for supervised learning-based illumination enhancement methods. Unsupervised illumination enhancement methods tend to cause color shifts. Since vehicle color is an important feature, color shift reduces model recognition capability. Moreover, most methods cannot adapt to different degrees of illumination, greatly reducing the practical value of vehicle re-identification tasks. To address these issues, we propose a novel self-calibrated module that improves image illumination. The module can be directly installed in front of the re-identification model and jointly trained with it, without the need for normal lighting images or additional loss functions, to achieve image illumination enhancement and not affect the original vehicle color. The module can adapt to different degrees of illumination during both daytime and nighttime, without considering the illumination distribution of the original training data. This paper proposes a novel vehicle re-identification dataset that includes training and testing datasets for both daytime and nighttime scenarios. Through experiments, we verify that the proposed method effectively improves the illumination of images and enhances the performance of vehicle re-identification networks under low-light conditions.