An intelligent food waste identification and analysis system based on convolution neural network
Xiyuan Sun, Yanxi Li, Zhimei Ouyang, Jinyuan Ouyang, Guangjing Yang, Enchang Sun · 2024
To achieve sustainable development, an efficient and time-effective management of waste, including household, industrial, and food waste, is crucial. This paper introduces an intelligent food waste (FW) identification and analysis system based on convolution neural network (CNN), significantly enhanced by the application of fine-tuning. This approach involves modifying the CNN model while retaining some original structures and pre-trained parameters, and training it using a new dataset. The implementation of this method has led to a remarkable increase in the accuracy of the CNN model from 94.7% to 97.6%, a reduction in the training time cost by approximately 82.3%, and a decrease in the number of parameters that need to be trained by about 96.3%. The proposed system, an Internet of Things (IoT) system, comprises a sensing layer, network layer, data storage layer, and application layer. It autonomously identifies the type of FW using CNN, collecting and analysing food waste data, including date, weight, type, and reason. The lightweight neural network MobileNetV2 was employed because of its low computing requirements. To further enhance accuracy and reduce resource costs, fine-tuning technology was applied using a new training dataset and a new output layer in the neural network. This paper presents a comparative analysis of the original and the improved CNN model, demonstrating the significant improvements achieved through the application of fine-tuning. These results suggest that fine-tuning can enhance the accuracy and training efficiency of CNN by saving time costs and parameters, thereby contributing to the application of the CNN and fine-tuning.