Weakly supervised object detection aims to use only image-level class labels to train the detector, which solves the problem of difficult data label acquisition and is of great research significance. Existing methods in weakly supervised object detection (WSOD) typically decompose the image detection task into a proposal (patch) classification problem. Due to the lack of an effective region proposal generation strategy and bounding box supervision, WSOD often highlights only the most discriminative parts of an object. While some approaches improve region proposals or suggest post-processing strategies to refine detection results, they usually fail to bridge the gap between classification and localization effectively in terms of feature representation. To address this limitation, we propose an Integrality Weakly Supervised Object Detection model (In-WSOD). In our model, we proposed a trainable Weakly Supervised Proposal Generation Network (WSPGN) that provides both enhanced region proposals and pseudo-bounding box labels. We further propose an Object Integrality Attention (OIA) to make the extracted features with classification and localization consistency. For adjusting the coordinates of output bounding boxes, we introduce a regression branch to make our In-WSOD detect objects more integrally. We conduct extensive experiments on PASCAL VOC 2007, VOC 2012 and MS-COCO 2017 datasets, where the proposed In-WSOD consistently outperforms other SOTA WSOD methods. Ablation studies further demonstrate the effectiveness of our WSPGN and OIA.