Improving Robot Perception Skills Using a Fast Image-Labelling Method with Minimal Human Intervention

Abstract

Abstract: Robot perception skills contribute to natural interfaces that enhance human–robot inter-actions. This can be notably improved by using convolutional neural networks. To train a convolutional neural network, the labelling process is the crucial first stage, in which image objects are marked with rectangles or masks. There are many image-labelling tools, but all require human interaction to achieve good results. Manual image labelling with rectangles or masks is laborintensive and unappealing work, which can take months to complete, making the labelling task tedious and lengthy. This paper proposes a fast method to create labelled images with minimal human intervention, which is tested with a robot perception task. Images of objects taken with specific backgrounds are quickly and accurately labelled with rectangles or masks. In a second step, detected objects can be synthesized with different backgrounds to improve the training capabilities of the image set. Experimental results show the effectiveness of this method with an example of human–robot inter-action using hand fingers. This labelling method generates a database to train convolutional networks to detect hand fingers easily with minimal labelling work. This labelling method can be applied to new image sets or used to add new samples to existing labelled image sets of any application. This proposed method improves the labelling process noticeably and reduces the time required to start the training process of a convolutional neural network model.