AI-Driven Machine Vision for Automated Defect Classification

Abstract

Visual inspection is a critical part of modern food production, where product quality and safety must be ensured at high speed and with high reliability. Manual inspection is time-consuming and prone to human error, which motivates the use of automated machine vision systems. The objective for this project was to investigate whether a camera-based system combined with deep learning could be a suitable solution for inspecting vacuum-sealed sausage packages. This was done by designing and implementing a prototype inspection system. The final system consists of multiple industrial cameras connected to an embedded computing platform, where images of the packages are captured and processed. With these multiple images create a overview of the product . A convolutional neural network is used to analyze the images in order to assess if the package is defect. The network is trained on a dataset collected during the project and is designed to perform the inspection automatically in real time. The results show promising performance for the use of convolutional neural networks in this type of industrial inspection task, as the system is able to detect relevant visual features and handle variations in lighting and positioning. However, there is still room for improvement, particularly in terms of dataset size, camera setup, and further optimization of the network architecture. This project concludes that a camera-based inspection system using deep learning is a viable solution for automated quality control of packaged food products in an industrial environment.