다양한 크기의 식재료 분할을 위한 Mask R-CNN 기반의 확장된 인공 데이터셋 생성

Abstract

This work proposes a scale-invariant instance segmentation method for images acquired from a real-time camera. It is challenging to detect and segment an exact shape by removing background (named as an instance) of a deformable semi-solid object such as food materials. In this work, we consider the segmentation with the cases of various sizes of an object and multiple objects overlapped with each other. To do this, we address an augmented dataset generation method, which extends dataset from small number of base objects, a fundamental dataset. Our method is based upon data augmentation, which is well known that it is an effective way to improve the segmentation performance. Our method addresses the generation of dataset with various scales using small number of original dataset. It is relatively simple in method but provides better performance. We also propose how to choose a target object (food material) with its centroid for grasping. Through diverse experiments using real-time images, we demonstrate that the proposed algorithm segments scale-invariant object masks and is successfully implemented for a robotic hand to grasp a food material. It is also compared with the state-of-the-art segmentation algorithm. As a result, the proposed method shows 74%, 85%, and 78% in accuracy, recall, and precision while the original dataset shows 67%, 79%, and 70%, respectively.