Fast Online Incremental Attribute-based Object Classification using Stochastic Gradient Descent and Self-Organizing Incremental Neural Network

摘要

In computer image processing, traditional object classification methods focus on visual objects such as cars, dogs, and airplanes. However, these objects are not easy to scale, require extensive training for all classes, and are susceptible to unseen object classification failure. Recently, a new object classification approach called attribute-based object classification has been introduced by considering visual adjectives that are easily discernible, such as "red, furry, has legs", instead of visual objects. Based on this concept, knowledge of an unseen object's categories can be conveyed by a description of its attributes. In this study, we proposed the first online incremental learning system for attribute-based object classification. Our proposed method can overcome online incremental learning tasks, and the learning and classification process is very fast. The presented experimental results are the first reported results of online incremental attribute-based object classification and hence can serve as a reference with which other results are compared. Moreover, our computation time is also more than 90% faster than other offline approaches.