샴 네트워크 기반의 적은 데이터셋을 이용한 유사한 형태의 특정 객체 인식 연구

Abstract

Target object recognition by a mobile robot is based on deep learning. To train a deep learning model to differentiate between target and non-target objects, a sufficiently large dataset that can extract features for classifying objects is required. In a general working environment, securing a sufficiently large dataset is difficult because the number of people designated as workers is constantly changing. Classifying the shapes of objects that are similar requires larger datasets than classifying the shapes of objects that are not similar. Therefore, a method for recognizing target objects with similar shapes using a small dataset is required. This paper proposes a Siamese-network–based target object recognition method for recognizing objects with similar shapes based on a small dataset. A trained Siamese network is used to recognize whether the input object is the target object based on its similarity to the target object. The results of target object recognition using the proposed method were experimentally analyzed. Further, ResNet-50 was used to evaluate the performance of the proposed method. Our findings show that the proposed method recognized objects with a difference of approximately 6% using a small dataset, indicating its higher efficiency than the classification-based object recognition method.