Answers are grounded in this paper and cite their sources.
To address the critical challenges in cotton maturity detection—such as small target recognition difficulties, complex background interference, variable lighting conditions, and severe occlusions caused by leaves and branches that lead to missed and false detections—this study proposes BTDM-DETR, an enhanced real-time detection model based on RT-DETR. To mitigate the loss of fine-grained features caused by convolutional operations and significantly decrease computational and memory access requirements, we introduce the FBRG-Block into the backbone network. Crucially, the standard AIFI network structure in the Transformer encoder is replaced with the proposed Local-Enhanced Token Statistics Self-Attention (LE-TSSA) module. By integrating a local spatial prior and dynamic channel recalibration, LE-TSSA preserves high-frequency local details essential for small targets while achieving linear computational efficiency. Furthermore, we design a Multi-Scale Adaptive Gated Fusion (MSAGF) module to adaptively combine local and global representations, alongside a Dynamic Pooling Cooperative Attention (DPSA) module that strengthens feature discriminability against intricate background interference.The effectiveness and practical feasibility of the proposed framework were rigorously established through comprehensive training and validation on a novel, self-collected private dataset manually captured from real-world cotton fields in Xinjiang, which captures authentic environmental complexities.Experimental results on the self-collected "cotton_maturity" dataset demonstrate that BTDM-DETR significantly outperforms the original RT-DETR-R18 in complex field environments, achieving absolute increases of 3.68% in precision, 3.14% in recall, 3.73% in mAP@0.5, and 6.05% in mAP@0.5:0.95.Moreover, to validate its practical engineering applicability, the model was deployed on an NVIDIA Jetson Nano edge device using TensorRT optimization, achieving a real-time inference speed of 23.5 FPS. A PyQt5-based visualization system was also developed to enable the real-time monitoring of live video streams. Generalization experiments on additional public datasets further validate the model's robustness. This advancement not only pushes the boundaries of target detection in unstructured environments but also provides a highly efficient, deployable solution for automated agricultural harvesting systems.
No takes yet. Share an insight, caveat, or question.
Li et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: