TL;DR
This paper introduces SP-YOLO, a novel deep learning model designed for real-time, accurate detection of pests in sugar beet fields. By integrating a hybrid CNN-Transformer backbone, a multi-scale feature fusion neck, and a lightweight depthwise separable convolution block, SP-YOLO significantly improves detection accuracy and speed, outperforming existing methods on a custom sugar beet pest dataset and demonstrating robustness on other datasets.
What Problem Does This Paper Address?
Sugar beet is a vital crop, but its yield and quality are severely threatened by pest infestations. Traditional pest detection relies on manual inspection, which is time-consuming, labor-intensive, and prone to human error. While computer vision and deep learning have shown promise, existing methods face challenges in agricultural settings: pests often blend into complex backgrounds, exhibit significant scale variations, and require models that are both accurate and efficient enough for real-time deployment on edge devices. Current models often struggle with these issues, leading to missed detections or slow processing.
Methods
The core of this work is the SP-YOLO model, built upon the YOLO11 architecture. Key innovations include:
- Hybrid CNN-Transformer Backbone (CAT): To address the limitations of traditional CNNs in capturing long-range dependencies, a novel CAT module is introduced. This module combines CNNs for local feature extraction with transformer blocks for global context, enhanced by a channel selection strategy to manage computational cost.
- Depthwise Separable Convolution Block (DSCB): This lightweight module is designed to extract multi-scale features and enlarge the receptive field, improving detection of pests at various sizes. It decomposes standard convolutions into depthwise and pointwise operations, significantly reducing parameters and computation.
- Cross-Layer Path Aggregation Network (CLPAN): To better fuse low-level and high-level features for multi-scale detection, the standard FPN and PAN structures are enhanced into CLPAN. This mechanism preserves fine-grained details while integrating semantic context, crucial for differentiating pests across scales.
Datasets
- BeetPest Dataset: A new, custom dataset was created for this study, specifically for sugar beet pests. It comprises 5318 images across 12 common pest species and 1 type of pest trace. The dataset was sourced from real field photographs, internet images, and the publicly available IP102 dataset, with a focus on diversity in pest species, scales, and complex backgrounds.
- IP102 Dataset: A subset of this large-scale insect pest dataset was used to supplement the BeetPest dataset, with careful selection and manual review to address labeling issues and class imbalance.
- RTB Dataset: This dataset, containing images of six common invasive pine tree pests with high morphological similarity, was used to evaluate the model's robustness and generality.
Key Insights and Ideas
- Hybrid Architectures are Key: Combining the strengths of CNNs (local features) and Transformers (global context) is crucial for robust pest detection in complex environments.
- Multi-Scale Feature Fusion is Essential: Agricultural pests vary greatly in size. Effectively fusing features from different network layers is vital for capturing both small and large targets.
- Lightweight Modules for Efficiency: Techniques like Depthwise Separable Convolutions are critical for developing models that are accurate enough for real-world applications but also efficient for deployment on resource-constrained edge devices.
- Custom Datasets Matter: Existing datasets may not fully capture the nuances of specific agricultural scenarios (e.g., complex backgrounds, specific pest species). Creating tailored datasets is essential for optimal performance.
Key Results
SP-YOLO demonstrated superior performance compared to several state-of-the-art object detection models on the BeetPest dataset:
- Accuracy: Achieved a mAP@50 of 0.884 and mAP@50:95 of 0.612, outperforming the baseline YOLO11n (0.835 mAP@50, 0.596 mAP@50:95) and other advanced models.
- Precision and Recall: Reported a precision (P) of 0.887 and recall (R) of 0.831, also showing improvements over YOLO11n.
- Speed: Achieved a detection speed of 136 FPS, meeting real-time requirements.
- Efficiency: While slightly increasing parameters (8.5M vs. 2.6M for YOLO11n), the model maintains a manageable computational cost and significantly outperforms heavier models like Faster R-CNN.
- Robustness: Showed excellent performance on the RTB dataset, highlighting its versatility.
- Ablation Studies: Confirmed the effectiveness of each proposed module (CAT, DSCB, CLPAN), with the full SP-YOLO model achieving the best results.
Limitations and Future Work
- Class Imbalance: The BeetPest dataset exhibits class imbalance (e.g., 770 images for Agrotis vs. 114 for "traces of damage"), which can bias training. While data augmentation was used, more robust strategies are needed.
- Pest Damage Traces: While the model showed promise in identifying damage traces, expanding this to multiple pest species and integrating it with pest location identification could further aid farmers.
- Multimodal Integration: The authors suggest integrating image data with textual descriptions of pests to create a multimodal model, which could capture more complex variations and have greater practical significance.
- Further Research: Future work should focus on developing more effective strategies for addressing class imbalance and exploring multimodal approaches for enhanced pest detection.
Summary updated