Liu Shacheng
Hunan Institute of Science and Technology, Yueyang, China.
Front Neurorobot. 2023 Jan 11;16:1083835. doi: 10.3389/fnbot.2022.1083835. eCollection 2022.
The package design can transmit the value of a product to consumers visually and can therefore influence the consumers' decisions. The traditional package design is an open-loop process in which a design can only be evaluated after the products are sent to the market. Thus, the designers cannot refine the design without any helpful advice. In this paper, a robust robot image classification is proposed to help the designers to evaluate their package design and improve their design in a closed-loop process, which is essentially the establishment of a cyber-physical system for the package design. The robust robot image classification adopts the total variation regularization, which ensures that the proposed robot image classification can give the right answers even if it is trained by noisy labels. The robustness against noisy labels is emphasized here since the historical data set of package design evaluations may have some false labels that can be equivalently regarded as disturbed labels from the true labels by noises. To validate the effectiveness of the proposed robot image classification method, experimental data-based validations have been implemented. The results show that the proposed method exhibits much better accuracy in classification compared to the traditional training method when noisy labels are used for the training process.
包装设计能够以视觉方式向消费者传递产品价值，从而影响消费者的决策。传统的包装设计是一个开环过程，在这个过程中，只有在产品投放市场后才能对设计进行评估。因此，设计师在没有任何有用建议的情况下无法改进设计。本文提出了一种鲁棒的机器人图像分类方法，以帮助设计师在闭环过程中评估他们的包装设计并改进设计，这本质上是为包装设计建立一个信息物理系统。鲁棒的机器人图像分类采用全变差正则化，这确保了所提出的机器人图像分类即使在使用有噪声标签进行训练时也能给出正确答案。这里强调对有噪声标签的鲁棒性，是因为包装设计评估的历史数据集可能有一些错误标签，这些错误标签可以等效地被视为被噪声干扰的真实标签。为了验证所提出的机器人图像分类方法的有效性，已经进行了基于实验数据的验证。结果表明，当在训练过程中使用有噪声标签时，与传统训练方法相比，所提出的方法在分类方面表现出更高的准确率。
