<p>The weakly supervised object localization task uses image-level labels to train object localization models. Traditional convolutional neural network (CNN)-based methods usually localize objects using a class activation map. However, the class activation map usually suffers from the problem of activating a small part of the object that is most discriminative. Meanwhile, the methods based on the Vision Transformer can capture long-range feature dependencies but tend to ignore local feature details. In this paper, we innovatively propose a dual-branch contrastive learning (DBC) method that consists of a Transformer and a CNN branch. The method can effectively separate the background and foreground of an image and fuse the features of Transformer and CNN through contrastive learning. Specifically, the method separates the background and foreground representations of the image using the initially generated class-agnostic activation maps. Then, the representations of the same image from different branches form positive pairs for contrastive learning. The background and foreground representations from the same branch form negative pairs. Finally, the DBC method forces the model to separate the background and foreground representations through negative contrastive loss and makes the model fuse the features of two branches through positive contrastive loss. Experiments on the ILSVRC benchmark show that the proposed method can achieve a Top-1 localization accuracy of 59.9% and a GT-known localization accuracy of 71.7%, which are better metrics than those of the state-of-the-art methods with the same parameter complexity.</p>