This is published under the terms of CC BY-NC.
Downloaded: 308 times
It is known that various implicit bias occur in Neural Networks due to their structural restrictions. Among them, texture bias caused by the convolution of CNNs has a significant impact on recognition performance. This paper shows that models with strong texture bias degrade recognition performance on datasets with large shape features, and to compensate for this characteristic of CNNs we introduce a method to increase their bias toward shapes rather than textures. Our method uses a simple image decomposition technique to create a shape-dominant dataset and then build a model with shape bias using the dataset. We experimentally show that the network can be biased towards shape without a significant loss of recognition accuracy compared to CNNs trained using conventional ImageNet. Additionally, we demonstrate that the CNN built by the proposed method obtains a higher recognition accuracy for shape-dominant images than those created using conventional methods.