SP-DETR: Superior point weak semi-supervised DETR with teacher–student paradigm for crop and weed detection

Recent years have witnessed extensive applications of deep learning-based computer vision techniques in agriculture for efficient digital management. Among these applications, field crop detection has emerged as a key research focus in smart agriculture. While most detection algorithms rely on supervised learning and have achieved remarkable success with large-scale annotated datasets, the annotation process remains prohibitively expensive. As an efficient alternative, point annotations present significant advantages over bounding box and image-level annotations in terms of time efficiency. However, it remains an unresolved challenge to effectively utilise point annotations to enhance semi-supervised object detection. In this work, a Superior Point weakly semi-supervised DETR (SP-DETR) with teacher–student paradigm is presented for crop and weed detection. Specifically, a Point Amplification Module (PAM) is first proposed to connect point annotations to objects in many-to-one matching way. Next, dual parallel branches are equipped with point encoder to align the point annotation and input image semantically. Then, like-reference points and decoder queries are integrated into Denoised Decoder (DD) to establish stable Hungarian matching with fast convergence. Finally, Colour-guided Point Annotation (CPA) is employed to ensure accurate target identification against background interference. Experimental results demonstrate the effectiveness of SP-DETR on BWD dataset when compared with other existing state-of-the-art approaches. In addition, ablation studies are carried out to verify the contribution of each proposed component. The source code is released to the public at https://github.com/731120464/SPDETR . • SP-DETR detects crop and weed in weakly semi-supervised DETR with teacher–student architecture. • The encoder and decoder are optimised with Point Amplification Module and dual parallel branches. • SP-DETR achieves superior detection performance on BWD dataset.