Transformer-based smart inspection for agricultural products via x-ray images
Chaoyu Dong, Alvin Wei Wen Tan, Tong Liu, Fang Cheng, Qian Kemao · 2024
The supply chain of agricultural products is intricately linked to the daily lives of people. In light of rising import and export quantities, the need for a prompt and efficient inspection system has become increasingly pressing. Without opening baskets and manually sorting, a smart inspection scheme is designed in this work leveraging X-ray images and transformer neural network. Due to its penetrating capabilities, X-ray enables a direct examination of agricultural products within a basket, a task that normal vision devices are unable to accomplish. Taking into account the varying shapes and combinations of agricultural products, we introduce a transformer-based deep neural network for type identification. Additionally, a dataset augmentation process is developed inspired by computed tomography generating 1,6000 X-ray images. Through experiments, the proposed smart inspection scheme is proven to be feasible and works efficiently. The inspection accuracy for both single-type and mixed-type agricultural products on the established dataset exceeds 90%.