TL;DR
This paper introduces YOLO11-ARAF, a novel deep learning model for accurate and efficient apple detection in complex orchard environments. By incorporating a new CARConv module for rotational invariance and an AFGCAM attention mechanism for challenging lighting, the model achieves improved detection performance. Further optimization through knowledge distillation results in a lightweight YOLO11n version suitable for edge devices, doubling inference speed while maintaining high accuracy.
What Problem Does This Paper Address?
Automated apple harvesting relies heavily on accurate fruit detection. However, real-world orchards present significant challenges: varying illumination (dazzling sun, shadows, low light), complex backgrounds with branches and leaves, and diverse apple appearances (color, size, occlusion). Existing object detection models, even advanced ones like YOLO variants, struggle with these conditions, leading to insufficient recognition performance and poor generalization. This limits the practical deployment of autonomous apple-picking robots. The authors aim to develop a model that is both highly accurate and computationally efficient, making it suitable for deployment on resource-constrained edge devices commonly used in agricultural robotics.
Methods
The core of the proposed solution is an enhanced YOLO11 architecture, named YOLO11-ARAF. The key innovations are:
- CARConv Module: This module replaces the standard C3k2 bottleneck in the YOLO11 backbone. It integrates Adaptive Rotational Convolution (ARConv), which allows the convolutional kernel to dynamically adapt to the orientation of objects. This is crucial for apples, which can appear at various angles due to natural growth and camera perspectives.
- AFGCAM Attention Module: An enhanced version of the AFGCA attention mechanism, AFGCAM is incorporated into both the backbone and neck of the YOLO11 model. It uses a dual-pooling strategy (Global Average Pooling and Global Max Pooling) to capture both broad contextual features and fine-grained local details. This helps the model better discern apples under challenging lighting and occlusion.
- Knowledge Distillation: To address the increased computational cost of the enhanced YOLO11-ARAF model, knowledge distillation is employed. A larger, more accurate "teacher" model (YOLO11-ARAF) is used to train a smaller, faster "student" model (YOLO11n). This process transfers the learned knowledge, enabling the student model to achieve high accuracy with significantly reduced parameters and improved inference speed.
Datasets
- Custom Apple Dataset: The authors constructed a new dataset consisting of 3942 annotated images of apples. This dataset was specifically designed to capture diverse real-world orchard conditions, including varied illumination (normal, dazzling, night, complex, overexposure, backlighting), different shooting angles, and varying shading levels. This diverse dataset is crucial for training and validating the model's robustness.
- Existing Datasets: While not explicitly detailed for training, the paper mentions comparative experiments against other YOLO series models and RTDETR, implying these models were likely evaluated on standard benchmarks or the authors' custom dataset.
Key Insights and Ideas
- Rotational Invariance is Key: The success of CARConv highlights that standard convolutional layers can struggle with objects at varying orientations. Explicitly addressing rotational adaptivity is vital for robust detection in natural scenes.
- Dual Attention for Robustness: Combining global and local feature extraction via AFGCAM's dual-pooling strategy proves more effective than single-pooling methods for handling complex visual challenges like lighting variations and occlusions.
- Distillation for Practicality: Knowledge distillation is a powerful tool for bridging the gap between high-accuracy research models and deployable, efficient edge-device solutions. The YOLO11n model demonstrates that significant speedups are achievable without substantial accuracy loss.
- Dataset Diversity Matters: The creation of a custom dataset reflecting diverse orchard conditions was essential for training a model that generalizes well beyond controlled environments.
Key Results
- YOLO11-ARAF Performance: The improved YOLO11-ARAF model achieved 89.4% accuracy, 86% recall, 92.3% mAP@50, and 64.4% mAP@50:95. These metrics represent improvements of 0.3%, 1.1%, 0.72%, and 2% respectively, over the baseline YOLO11 model.
- Lightweight YOLO11n: The distilled YOLO11n model significantly reduced parameters and doubled the inference speed (FPS) compared to the improved YOLO11-ARAF, making it suitable for edge deployment.
- Robustness in Challenging Scenarios: The model demonstrated superior performance on challenging sub-validation sets, particularly in complex lighting conditions, showing a 0.7% and 2% accuracy improvement over the baseline in complex background and complex lighting scenarios, respectively.
- Comparative Performance: YOLO11-ARAF outperformed other YOLO series models and RTDETR variants in key metrics, especially mAP@50:95, indicating better localization accuracy.
Limitations and Future Work
- Dataset Specificity: The current dataset primarily features a single apple variety. The authors acknowledge that this might limit the model's generalization to other apple varieties with different visual characteristics. Future work should involve collecting diverse apple samples to retrain the model.
- Model Size: While knowledge distillation improved efficiency, the YOLO11-ARAF model's parameter size is still larger than some typical convolutional models.
- Future Directions: The authors suggest exploring model pruning techniques to eliminate redundant layers and investigating more lightweight distillation student models and algorithms for further efficiency gains. They also note that complex background conditions (severe occlusion) pose a greater challenge than complex illumination, suggesting this as a potential area for future optimization.
Summary updated