Training object detectors with only image-level annotations is an important problem with a variety of applications. However, due to the deformable nature of objects, a target object delineated by a bounding box always includes irrelevant context and occlusions, which causes large intra-class object variations and ambiguity in object-background distinction. For this reason, identifying the object of interest from a substantial amount of cluttered backgrounds is very challenging. In this paper, we propose a decoupled attention-based deep model to optimize region-based object representation. Different from existing approaches posing object representation in a single-tower model, our proposed network decouples object representation into two separate modules, i.e., image representation and attention localization. The image representation module captures content-based semantic representation, while the attention localization module regresses an attention map which simultaneously highlights the locations of the discriminative object parts and down weights the irrelevant backgrounds presented in the image. The combined representation alleviates the impact from the noisy context and occlusions inside an object bounding box. As a result, object-background ambiguity can be largely reduced and background regions can be suppressed effectively. In addition, the proposed object representation model can be seamlessly integrated into a state-of-the-art weakly supervised detection framework, and the entire model can be trained end-to-end. We extensively evaluate the detection performance on the PASCAL VOC 2007, VOC 2010 and VOC2012 datasets. Experimental results demonstrate that our approach effectively improves weakly supervised object detection.