Kidney Bean Detection in Complex Agricultural Scenarios Based on the KidneyB-YOLO Model

Jianjia QiChunxiang Liu^*Hongxia Chu

• Heilongjiang Institute of Technology, Harbin, China

This study addresses the problem of detecting multiple targets and occluded soybeans in the complex agricultural scenarios under the open-field trellis cultivation, and proposes a target detection model based on kidneyB-YOLO. The model integrated the dynamic con-volution module, the Deformable Attention Transformer (DAT) attention mechanism, the feature fusion detection head improved Adaptively Spatial Feature Fusion (ASFF) detection method, and the Focaler-SIoU loss function, enhancing the accuracy and robustness of bean pod detection. The dynamic convolution module adaptively adjusted parameters based on the input, reducing computational overhead and improving the model's ability to represent complex scenes and small targets. The dynamic convolution module, when added to the neck network, achieved the best accuracy improvement. The DAT attention mechanism introduced a deformable attention mechanism, focusing only on a small key region of the image, which was suitable for detecting long, small bean pod targets. The feature fusion detection head ASFF, an adaptive spatial feature fusion method, filtered out conflicting target information, enhancing scale invariance and improving detection accuracy. The new loss function module, Focaler-SIoU, was formed by combining the SIoU loss function with the Focaler-IoU algorithm, giving extra attention to the specific features of different samples and addressing the issue of inconsistent sample resolution. An open-canopy environment bean pod dataset was constructed for model training and validation. The results showed that the KidneyB-YOLO model exhibited significant improvements compared to the original YOLOv8n, especially in occluded scenarios. The model achieved a detection performance of 85.90% mAP with a computational complexity of 12.4 GFLOPs, a model size of 12 MB, and operated at 32.5 FPS, demonstrating strong robustness and generalization capability in the fruit detection task for kidney bean harvesting under open-air trellises.