Non-strict Attentional Region Annotation to Improve Image Classification Accuracy

Although convolutional neural networks (CNNs) have been widely used in image classification, a large amount of training data is required for training models. Especially in industrial applications, not only is there a demand for reducing the cost of labeling images, but also, it is difficult to obtain a large amount of training images. Several approaches have been proposed to improve classification accuracy by embedding human knowledge other than labels into the models. However, in these approaches, the annotation methods tend to become complex and applicable only to specific model architectures corresponding to those methods.In this study, we propose a new approach to improve the image classification accuracy of CNNs by adding human knowledge as simple region annotations. In our method, annotators draw the regions that attract the most attention as a single ellipse per image for classifying image categories. This annotation is called non-strict attentional region annotation (NARA) and is used for training CNNs together with the labels. We also propose a training method using NARA, which is applicable to common CNNs. For the proof of concept, we have added region annotations to all 5,000 images of the STL-10 training dataset. Through experiments, we demonstrate that our approach is effective to improve image classification accuracy with relatively inexpensive additional cost. Our annotation dataset is available at: https://git.io/stl10nara.