A Semi-automatic Annotation System for Imperfect Grain Images

In the task of agricultural imperfect grain detection, there are several challenges, including time-consuming manual classification, inconsistent annotating standards among annotators, and difficulties in managing a large volume of images. To address some of these issues, a semi-automatic annotation system for imperfect grain images is presented in this paper. Images from field devices are stored in the system in the form of datasets, differentiating grains from various geographical regions. To address the limited computational resources of devices, the Mobile-DCL algorithm is proposed to achieve fine-grained classification, conserve computational resources, and accelerate inference speed simultaneously. To enhance the system's domain adaptability and expedite the model iteration speed, a semi-automatic model generating methodology, "model inference — expert validation — model fine-tuning", is employed in this paper. Test results demonstrate that the presented system enhances the level of automation in imperfect grain detection during agricultural production, and improves the annotation experience for relevant personnel.