Object detection for industrial food recognition

This work is dedicated to developing an object detection system for food recognition inside a professional oven. Deploying an object detector will automate the oven-user interaction, and decrease the mistake rate, associated with the wrong choice of cooking parameters. A solution must comply with numerous limitations, such as low-performance hardware, scarce data of 30 images per recipe, low ambient light, and constraints on the recognition speed. The result of the work is a prototype with a light setup, camera, and object detector, running in an embedded system. In addition, we cover data collection procedures, and techniques for increasing the model performance (e.g. data augmentation, zone assignment algorithm, unsure inference algorithm, confusion matrix evaluation, etc.). As a result, when trained on a plastic dataset of five recipes, the prototype correctly responds to 98% of validation images. This has accomplished the proof of concept for UNOX S.p.a. company.