This paper presents a novel deep learning model, PodNet, designed for accurate and efficient soybean pod counting and localization from high-resolution images. The model aims to address the challenges of pod occlusion, uneven pod distribution, and the need for lightweight and real-time applications. PodNet employs a lightweight encoder and an efficient decoder, leveraging a deep convolutional network with a feature pyramid network (FPN) to enhance information decoding. The model is evaluated using a high-resolution dataset of soybean pods, demonstrating superior performance compared to existing methods. Key contributions include: 1. **Model Architecture**: PodNet uses a lightweight encoder (CSPDarknet) and an effective decoder with ASFF (Adaptively Spatial Feature Fusion) and Mlt-ECA (Multi-Efficient Channel Attention) modules to improve feature fusion and context aggregation. 2. **Efficiency**: PodNet achieves an R² of 0.95 for pod count prediction with only 2.48M parameters, significantly fewer than the state-of-the-art model YOLO POD, while maintaining a much higher FPS. 3. **Real-Time Efficiency**: PodNet operates at a frame rate of 43.67 FPS on a GTX1080Ti GPU, making it suitable for real-time deployment on low-cost devices. The paper also discusses the challenges in plant counting, such as pod occlusion and uneven distribution, and provides a detailed analysis of the model's performance in various scenarios, including dense and complex backgrounds. Ablation studies validate the effectiveness of the proposed techniques. The authors conclude by highlighting the potential of PodNet for practical applications in high-throughput plant phenotyping and suggest future research directions to enhance the model's robustness and adaptability.This paper presents a novel deep learning model, PodNet, designed for accurate and efficient soybean pod counting and localization from high-resolution images. The model aims to address the challenges of pod occlusion, uneven pod distribution, and the need for lightweight and real-time applications. PodNet employs a lightweight encoder and an efficient decoder, leveraging a deep convolutional network with a feature pyramid network (FPN) to enhance information decoding. The model is evaluated using a high-resolution dataset of soybean pods, demonstrating superior performance compared to existing methods. Key contributions include: 1. **Model Architecture**: PodNet uses a lightweight encoder (CSPDarknet) and an effective decoder with ASFF (Adaptively Spatial Feature Fusion) and Mlt-ECA (Multi-Efficient Channel Attention) modules to improve feature fusion and context aggregation. 2. **Efficiency**: PodNet achieves an R² of 0.95 for pod count prediction with only 2.48M parameters, significantly fewer than the state-of-the-art model YOLO POD, while maintaining a much higher FPS. 3. **Real-Time Efficiency**: PodNet operates at a frame rate of 43.67 FPS on a GTX1080Ti GPU, making it suitable for real-time deployment on low-cost devices. The paper also discusses the challenges in plant counting, such as pod occlusion and uneven distribution, and provides a detailed analysis of the model's performance in various scenarios, including dense and complex backgrounds. Ablation studies validate the effectiveness of the proposed techniques. The authors conclude by highlighting the potential of PodNet for practical applications in high-throughput plant phenotyping and suggest future research directions to enhance the model's robustness and adaptability.