A Lightweight model for accurate soybean pod detection: YOLO-MobilePod

Soybean pods are critical indicators for evaluating soybean yield and quality. In natural environments, they exhibit substantial morphological variability and are predominantly small-scale objects, often characterized by dense distributions and frequent occlusions. Existing object detection algorithms still face considerable limitations in detecting small-scale and overlapping targets.

To address these challenges, this study proposes a lightweight and efficient soybean pod detection model, termed YOLO-MobilePod, which is developed based on YOLOv12. The proposed model adopts MobileNetV4 as the backbone network to reduce computational complexity while enhancing multi-scale feature extraction capability. To further improve the detection performance for small-scale soybean pods, YOLO-MobilePod introduces high-resolution feature fusion strategy in the neck network, thereby strengthening the representation of small-object features. In addition, dynamic convolution (DynamicConv) is incorporated into the regression branch of the detection head, where adaptive convolutional kernel combinations are employed to enhance feature modeling capacity while achieving additional model lightweighting. This study integrated a public dataset with a self-collected dataset and conducted both ablation and comparative experiments. The Experimental results demonstrate that, compared with the YOLOv12n model, YOLO-MobilePod achieves improvements of 2.07% in recall and 4.10% in mAP50-95, while reducing the number of parameters, FLOPs, and model size by 45.31%, 4.762%, and 38.18%, respectively. In addition, the inference speed increased by 8.768%.

These results demonstrate that the proposed model exhibits effective recognition performance for soybean pods with diverse morphologies and achieving improved detection accuracy while preserving model lightweighting, thereby providing a feasible technical reference for efficient soybean pod detection.