Grounding DINO for enhanced quarantine pest detection in grain imports

Le résumé fourni par la source

Multi-object detection, as one of the core tasks in computer vision, aims to identify all foreground objects in an image, determining their categories and locations. It has significant application prospects in fields such as automated detection, medical imaging, autonomous driving, security, and surveying. Currently, most pretrained large models have been open-sourced, which means we can treat them as general feature extractors for specific downstream tasks. Inspired by GLIP and DINO models, the Grounding DINO model was proposed, which can detect arbitrary objects based on text prompts with high detection accuracy. This project focuses on the problem of quarantine pests, such as invasive weeds and insects, in imported grain, which is crucial for quarantine inspection at ports. Grounding DINO was optimized and trained on a manually annotated dataset of grain weeds and insects, allowing the model to adapt and focus on the target domain, thereby enhancing its performance and potential application value in grain, weed, and insect recognition tasks. Experimental results show that the model exhibits excellent performance in object detection, achieving high average precision (AP [IoU=0.50:0.95]=0.867) and average recall (AR [IoU=0.50:0.95]=0.947), especially at an IoU of 0.50. Moreover, the comparison of pre- and post-training models proves the effectiveness of pre-training, leading to more prominent performance of the model in datasets of weeds, insects, and wheat images, thus improving detection accuracy.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.