Automatically marking object regions based on tagged images

Abstract

The performance of image classification can be greatly improved by suitable object proposals. The mainstream framework of object proposals (e.g. Faster R-CNN) needs to manually label every bounding box of each object on an image in the training or fine-tuning stage, which is a time consuming task. Thus, we propose an idea that object regions in each image can be generated firstly by a pre-trained Faster R-CNN based on other complete tagging data set (e.g. PASCAL VOC 2007). Then a fine-tuned convolutional neural network (CNN) on the current data set can be used to mark the object region automatically. Finally, these labeled object regions can be used to fine-tune the Faster R-CNN and CNN. In the animal classification data set of BOT 2016, experimental results show that our proposed method can greatly boost the average accuracy of image classification.