For the pixel level marked data of image segmentation is time consuming and expensive, the image segmentation under the weak supervised model with bounding box annotations was studied. The high-resolution pixel features were extracted by a pyramid densely sampled fully convolution network, and then the weakly supervised data by the GrabCut algorithm was transformed. Finally, the image features and the labeled data were jointly trained. A model of weakly supervised image semantic segmentation based on fully convolution network was constructed and verified on public data set. Experiments show that the constructed weak supervised model has a better segmentation effect compared with DET3-Proposed model, expectation-maximization model and Bbox-Seg model.