| Summary: | The grading system for raw unclean EBN plays a vital role in determining the market price between the EBN industry and swiftlet farming. The system also acts as a primary process to monitor the quality of EBN in the production line. However, the human visual system is subjective and based on the workers' experience, hindering a high performance in the grading system. Although the machine learning classifiers such as ANFIS and KMBA were more standardized and accurate, they required experience workers with the specific operation technique for the application. Therefore, a deep learning model with the self-learning ability on the feature extraction process and low human intervention was developed to solve the drawbacks of the human visual system and conventional algorithms. The transfer learning approach could save more computational power via a pre-trained model than build the model from scratch. It also reduces the labour-intensive and time-consuming issues in collecting the vast dataset to train the model. As a result, the best-fine-tuned model was ResNet50, with the highest accuracy of 92.51% among the five pre-trained models selected in identifying 13 of the EBN grades. The performance of the fine-tuned model outperformed the conventional classifiers of ANFIS (88.24%) and KMBA (85.60%) in the EBN grading system. Neuron activation and Grad-CAM analyses were proposed for visualizing the model's prediction on the EBN grades. The investigations aim to provide strong evidence that the fine-tuned model had learned the distinctive and relevant features for predicting the EBN grades. The EBN samples also fed into the deep dream images to enhance the features had detected by the model to indicate the respective EBN grades. The methods provide a better understanding to humans in the model's prediction for increasing the trustability of the model in the automatic EBN grading system.
|
|---|