BeanpodNet: A Dense, Occlusion-Aware Soybean Pod Detection Model

Counting mature soybean pods in field environments is a crucial challenge in soybean breeding. Although deep learning detectors, notably single-stage YOLO series, have advanced in agriculture, pod detection still struggles in complex environments with occlusion, dense pods, varied orientations, and large aspect ratios. To address these challenges, we propose BeanpodNet, a real-time soybean pod detection model based on a field rail-mounted high-throughput phenotyping platform and a single-stage detector. It integrates several innovative techniques to enhance detection efficiency and accuracy. First, a Super Token Sampling Vision Transformer (SViT) is employed to model global context and improve feature representation. Second, a novel multi-level attention mechanism (MultiSEAM) is introduced to enhance the detection of overlapping targets caused by occlusion. Lastly, we propose the Inner-IoU strategy for bounding box regression, accelerating the regression process for overlapping regions, improving both speed and accuracy. To enable precise soybean pod detection, we constructed the Podset dataset, comprising 1,200 field-collected images. Experimental results show that BeanpodNet achieves an [email protected] of 93.3%. Compared to state-of-the-art approaches, It demonstrates substantial improvements, highlighting its strong potential for practical applications.