Data Augmentation for Object Recognition of Dynamic Learning Robot

The training of deep learning networks for robot object recognition requires a large database of training images for satisfactory performance. The term “dynamic learning” in this paper refers to the ability of a robot to learn new features under offline conditions by observing its surrounding objects. A training framework for robots to achieve object recognition with satisfactory performance under offline training conditions is proposed. A coarse but fast method of object saliency detection is developed to facilitate raw image collection. Additionally, a training scheme referred to as a Dynamic Artificial Database (DAD) is proposed to tackle the problem of overfitting when training neural networks without validation data.