[4B1] Real-time deep learning-based vision inspection system for detecting and classifying defects in tomatoes
K Yang
TWI, UK
Supplier and packer companies are an essential link in the UK food supply chain and are dominated by manual labour due to the need for human handling and inspection skills. A prototype vision system with a camera, lenses, light source, enclosure, deep learning-based defect detector, graphical user interface and reject mechanism has been developed to ensure the fruit is clean and free from external visually detectable defects. In general, tomatoes are susceptible to a broad range of diseases and defects. The vision system has been designed to locate and classify 13 different defects in tomatoes: scarring, infection, splits, discolouration, contamination, mechanical damage, open stem scar, rot, shrivelled skin, pest exit holes, gold speckle, cat facing and rough skin. Additionally, the system can differentiate between a calyx scar (peduncular end) and true fruit defects. Via a commercial conveyor system, tomatoes will be delivered to the vision area in a single file, rotating on diablo rollers. The information captured from the vision system will be used, with any tomato failing to meet the visual quality specification rejected and diverted off the mainline.
