December 9, 2019 | Written by: Jian Xu, Guo Qiang Hu, and Jun Zhu
Categorized: AI | IBM Research-China
Share this post:
To train the model, the software had to analyze more than 30 types of impurities including insects, rocks and plastics. There is no room for error – so all impurities must be removed, including less harmful ones such as seeds and stalks, to improve the grade of the tea.
When you drink that cup of tea, you sure don’t want any fluff floating there. That’s the most critical step in tea production – removing impurities, be it stems, seeds, stones or even bugs. If they do make it into your drink, the brand may get fined; at the very least, you probably won’t buy from them again.
But spotting all the debris is tricky and time-consuming. Traditionally, it has been done manually, and still today, glove-clad inspectors carefully comb through raw tea leaves, straining their eyes. The process is tedious and difficult to scale.
To speed up the task and greatly improve accuracy, IBM AI scientists in China have developed an automated tea quality inspection line for a famous team company, based on a deep neural network detection model. The machine quickly and efficiently discarded impurities from raw leaves of Dian Hong – a fairly high-end Chinese black tea grown in Yunnan Province. The research was presented today at EmTech China.
To do that, first, a camera takes pictures of the leaves on a multi-stage production conveyor belt. The images are then sent to the AI model in the cloud, trained to detect impurities based on size and type. Once an impurity is spotted, a robotic arm whisks it away – and then does it all over again until the leaves are ready to make it into your favorite mug.
IBM Research – China team has filed 18 patents in object detection, defect classification, localization and segmentation and successfully built and applied the IBM cognitive visual inspection model.
To train the model, the software had to analyze more than 30 types of impurities including insects, rocks and plastics. There is no room for error – so all impurities must be removed, including less harmful ones such as seeds and stalks, to improve the grade of the tea. Stones, insets, plastics may cause serious food safety issues must be all removed, i.e. 0% skip rate for the AI model, but tea seeds and tea stalks, that doesn’t escalate to the level of food safety, still should be almost extracted out (<1% skip rate) to improve the quality grade of tea.
But analyzing images is not easy, as the quality of the photos taken for model training in the lab and those snapped at a real-world tea production facility varies greatly. That means that with the production environment images, the model performs much worse. Standard deep learning approaches require developers to collect and label enough training images from the production line, which is very labor-intensive. In the project, an advanced machine learning algorithm was applied, which can normalize the visual features under diverse environment, and detect the nuances between tea and impurities. The training requires only a small number of additional images taken from the production environment and saved considerable amount of model development effort and time.
To make the AI smarter, the scientists used a new, advanced model training process. They developed software able to suppress the learning of visual features in images taken in diverse environments while reinforcing the learning of the features that differentiate tea from all the debris. The training required a small number of additional images from the production environment, saving a lot of time while developing the model.
Next up for the research the researchers will adapt the inspection model of Dian Hong (tea type) to five additional tea types.
High Performance Visual Inspection Service Architecture – Squeezing the Most Out of Commodity Servers, 2018 IEEE International Conference on Web Services (ICWS), https://ieeexplore.ieee.org/document/8456382