부산대학교 도서관

Automated monitoring technology is necessary to ensure the safe consumption and management of meat. Here we propose a system for determining food freshness on the basis of data collected by embedded cameras and gas sensors using artificial intelligence of things (AIoT). Using the DeepLab V3+ model, images of meat obtained from embedded cameras underwent semantic segmentation into lean meat and fat, revealing the nature of rotting over time. Using convolutional neural network (CNN) learning, we classified meat into fresh, semi-fresh, and rotten and detected the transition among these. The Raspberry Pi server collected online images and gas sensor data, which were then processed by a mobile application. The sensor data were collected at 1-min intervals, and the means and standard deviations (SDs) of the valuable features were extracted. Meanwhile, the camera images were preprocessed, and their features were obtained using the transfer learning approach and MobileNet as the feature extractor. Next, a proposed 1-D-CNN (1D-CNN) was trained on the combined features for predicting different meat quality states. The accuracy of meat freshness detection using the proposed deep learning model reaches up to 99.44%. Our Internet-of-Things (IoT)-connected monitoring platform, including an embedded device, database server, and smartphone application, allows for quick and easy monitoring of meat quality. The experimental results confirmed the viability of our deep-learning-based meat-monitoring system with IoT applications in the real world.