Object-level benchmark for deep learning-based detection and classification of weed species

• DOI: 10.1016/j.cropro.2023.106561
• Publication Date: 2023-12-17T15:02:17Z
• Authors: A.S.M. Mahmudul H...; Dean Diepeveen; Hamid Laga; Michael G.K. Jones; Ferdous Sohel
• Keywords: Benchmark (surveying)

Weeds can decrease yields and the quality of crops. Detection, localisation, classification weeds in crops are crucial for developing efficient weed control management systems. Deep learning (DL) based object detection techniques have been applied various applications. However, such generally need appropriate datasets. Most available datasets only offer image-level annotation, i.e., each image is labelled with one species. practice, multiple (and crop) species and/or instances Consequently, lack instance-level annotations puts a constraint on applicability powerful DL techniques. In current research, we construct an dataset. The images sourced from publicly dataset, namely Corn It has 5997 plants four types weeds. We annotated dataset using bounding box around instance them crop or weed. Overall, contain about three average, while some over fifty boxes. To establish benchmark evaluated several models, including YOLOv7, YOLOv8 Faster-RCNN, to locate classify performance models was compared inference time accuracy. YOLOv7 its variant YOLOv7-tiny both achieved highest mean average precision (mAP) 88.50% 88.29% took 2.7 1.43 ms, respectively, image. YOLOv8m, YOLOv8, detected 2.2 ms mAP 87.75%. Data augmentation address class imbalance improves results 89.93% 89.39% YOLOv8. accuracy performed by this research indicate that these be used develop automatic field-level system.