利用人的分歧介入增强珍珠自动分拣可靠性研究

Research on Reliability Improvement for Pears Classification Based on Human Disagreement Intervention

摘要:

针对珍珠自动分拣应用场景,把优等珍珠分成次等品会导致严重利益损失，把有瑕疵珍珠分成上等品会造成产品质量争议和企业信誉受损；基于深度学习的人工智能技术依旧存在着难解释，鲁棒性差等缺点，导致难以实现更高精度的分拣技术。为衡量珍珠分拣准确度的现实需求和提升准确度在现有技术框架下的限制因素，研究提出了一种通过人的分歧介入提升分拣可靠性的方法。该方法引入两个独立AI系统用于珍珠分拣的预处理，然后通过二者之间的分歧引入人的介入干预，在较少的人力成本下达到了对机器算法可靠性的提升。定义了包括分歧准确指数和额外成本指数在内的性能评价指标，在公开珍珠数据集上，研究提出的方法以4.1%的额外人工成本提升了近4%的珍珠分拣精度，验证了方法的有效性。

Abstract:

According to the application scenario of pearl automatic sorting, dividing superior pearls into inferior ones will lead to serious loss of profits, and dividing defective pearls into superior ones will cause product quality disputes and enterprise reputation damage; Artificial intelligence technology based on deep learning is still difficult to explain and has poor robustness, which makes it difficult to achieve higher precision sorting technology. In order to measure the practical needs of pearl sorting accuracy and the limitations of improving accuracy under the existing technical framework, a method of improving sorting reliability through human disagreement intervention approach is proposed. This approach introduces two independent AI systems to pre-process pears classification, whose disagreements then bring human intervention. In such a way the reliability of machine algorithms is improved at relatively less human cost. The performance indices including disagreement precision index and additional cost index are defined and based on public dataset the proposed approach improves nearly 4% of the classification accuracy at only 4.1% increase of the human cost, which proves the effectiveness of the proposed approach.