Proceedings of the
9th International Conference of Asian Society for Precision Engineering and Nanotechnology (ASPEN2022)
15 – 18 November 2022, Singapore
doi:10.3850/978-981-18-6021-8_OR-08-0140
X-ray Artificial Intelligence Identification System for Agricultural Products
1Singapore Institute of Manufacturing Technology, A*STAR, 73 Nanyang Dr, 637662, Singapore
2School of Computer Science and Engineering, Nanyang Technological University, 50 Nanyang Ave, 639798, Singapore
3Advanced Remanufacturing and Technology Centre, A*STAR, 3 Cleantech Loop, 637143, Singapore
ABSTRACT
The growing trade of agricultural products is accompanied by the increasing challenges for product identification. Due to the mixture of various products in covered baskets or crates, conventional identification heavily relies on manual sorting, in which human operators must open the baskets and recognize internal products one by one. This conventional approach is not only time-consuming, but also suffering an accuracy issues in proper identification of the products, particularly with those with high similarity in appearance. To raise the identification efficiency and avoid careless mistakes, an artificial intelligence-assisted identification system is proposed and investigated. With this new method, the X-ray hardware collects vegetable images from different angles. Then a deep learning-based strategy is developed employing data augmentation, deep neural network, and feature visualization techniques, which formulates an end-to-end scheme for product identification. In the data augmentation stage, automatic type indexing is established mapping X-ray images to the encoded product types, while the collected raw X-ray images are compressed and randomly flipped to increase the processing speed and robustness. After that, an extensible network structure is proposed using the deep residual structure for the image feature extraction and parallel fully connected layers for the type identification. Besides, dropout layers are inserted to prevent the overfitting of the whole deep neural network. Finally, the principal component analysis is leveraged in the feature visualization part to enhance the result reliability and explainability. This integration of hardware and software techniques significantly releases human resources and achieves high detection accuracy, which notably improves the intelligence level of agricultural product identification. Detailed experiments are conducted with various types of agricultural products as well as different combinations. For mixed agricultural products in one basket, the minimum identification accuracy can reach 98.0% in the test dataset and the identification time for one X-ray image is only 0.18 seconds. The 1000-dimensional features extracted by the deep neural network are also visualized as well as the critical features from the convolutional kernel. For the first time, it is disclosed that root parts are core traits inside X-ray images intelligently learned for the type distinguishment.
Keywords: X-ray, Artificial intelligence, Agricultural product, Identification
1Singapore Institute of Manufacturing Technology, A*STAR, 73 Nanyang Dr, 637662, Singapore
2School of Computer Science and Engineering, Nanyang Technological University, 50 Nanyang Ave, 639798, Singapore
3Advanced Remanufacturing and Technology Centre, A*STAR, 3 Cleantech Loop, 637143, Singapore
ABSTRACT
The growing trade of agricultural products is accompanied by the increasing challenges for product identification. Due to the mixture of various products in covered baskets or crates, conventional identification heavily relies on manual sorting, in which human operators must open the baskets and recognize internal products one by one. This conventional approach is not only time-consuming, but also suffering an accuracy issues in proper identification of the products, particularly with those with high similarity in appearance. To raise the identification efficiency and avoid careless mistakes, an artificial intelligence-assisted identification system is proposed and investigated. With this new method, the X-ray hardware collects vegetable images from different angles. Then a deep learning-based strategy is developed employing data augmentation, deep neural network, and feature visualization techniques, which formulates an end-to-end scheme for product identification. In the data augmentation stage, automatic type indexing is established mapping X-ray images to the encoded product types, while the collected raw X-ray images are compressed and randomly flipped to increase the processing speed and robustness. After that, an extensible network structure is proposed using the deep residual structure for the image feature extraction and parallel fully connected layers for the type identification. Besides, dropout layers are inserted to prevent the overfitting of the whole deep neural network. Finally, the principal component analysis is leveraged in the feature visualization part to enhance the result reliability and explainability. This integration of hardware and software techniques significantly releases human resources and achieves high detection accuracy, which notably improves the intelligence level of agricultural product identification. Detailed experiments are conducted with various types of agricultural products as well as different combinations. For mixed agricultural products in one basket, the minimum identification accuracy can reach 98.0% in the test dataset and the identification time for one X-ray image is only 0.18 seconds. The 1000-dimensional features extracted by the deep neural network are also visualized as well as the critical features from the convolutional kernel. For the first time, it is disclosed that root parts are core traits inside X-ray images intelligently learned for the type distinguishment.
Keywords: X-ray, Artificial intelligence, Agricultural product, Identification
