In recent years, object detection has been applied to many fields. However, retraining using large amount of bounding-box labeled data is needed. This study improves the Faster RCNN method and solves the problem of detecting multi-object in images using few-shot single object training data without bounding-box annotation. We propose a non-classwise bounding-box regression layer, which is only trained by public dataset and used for product training image labeling and testing image detection. Combined with Grabcut method, a data augmentation method is proposed to generate multi-object product training image. The improved faster RCNN model is re-trained by these images. In addition, a re-identification layer is added to the Faster RCNN architecture and improves the detection performance.