Bread Browning Stage Classification Model using VGG-16 Transfer Learning and Fine-tuning with Small Training Dataset

Abstract

Convolutional neural network (CNN) is a popular tool to recognize image features even though its weakness requirement of massive training dataset. However, the implementation of the network in production process needs to worry about, that is, the deal with at least two constraints, small training dataset and the less diversity of browning stages among the bread production batches. This paper is aimed to achieve a high predictive accuracy model to classify the bread browning stage that is capable to deal with these constraints. With small training dataset of 900 original images from a production batch, the research performs five steps, starting with a few convolutional layers, adding image augmentation technique and transfer learning with pre-trained CNN model to enhance feature extraction with fine-tuning in final step. The final model of VGG-16 transfer learning and fine-tuning, trained with 18,000 artificial images, successfully achieves very high training accuracy of 98.89% and a very low loss of 2.86% at a small number of epochs 30 with its predictive accuracy of 100%.