Image annotation is a task of assigning semantic labels to an image. Recently, deep neural networks with visual attention have been utilized successfully in many computer vision tasks. In this paper, we show that conventional attention mechanism is easily misled by the salient class, i.e., the attended region always contains part of the image area describing the content of salient class at different attention iterations. To this end, we propose a novel attention shaping mechanism, which aims to maximize the non-overlapping area between consecutive attention processes by taking into account the history of previous attention vectors. Several weighting polices are studied to utilize the history information in different manners. In two benchmark datasets, i.e., PASCAL VOC2012 and MIRFlickr-25k, the average precision is improved by up to 10% in comparison with the state-of-the-art annotation methods.