Abstract
Accurate target detection in complex orchard environments is the basis for automatic picking and pollination. The characteristics of small, clustered and complex interference greatly increase the difficulty of detection. Toward this end, we explore a detector in the orchard and improve the detection ability of complex targets. Our model includes two core designs to make it suitable for reducing the risk of error detection due to small and camouflaged object features. Multi-scale texture enhancement design focuses on extracting and enhancing more distinguishable features for each level with multiple parallel branches. Our adaptive region-aware feature fusion module extracts the dependencies between locations and channels, potential cross-relations among different levels and multi-types information to build distinctive representations. By combining enhancement and fusion, experiments on various real-world datasets show that the proposed network can outperform previous state-of-the-art methods, especially for detection in complex conditions.
Introduction
Precision agriculture is moving from a futuristic dream to a field reality, but robotic systems still struggle with a fundamental human skill: seeing. In a dense orchard, a green kiwi fruit looks remarkably like a green leaf, and tiny flowers are often obscured by shadows or clustered together. Standard object detectors, such as the YOLO or R-CNN families, often stumble in these "messy" environments due to three primary hurdles: small target scales, camouflage where targets share patterns with the background, and complex interference from overlapping branches and shifting light.
A recent study published in Complex & Intelligent Systems (2024) by Wei Sun and colleagues introduces a specialized detector designed to overcome these challenges. By focusing on texture enhancement and smarter feature fusion, the team has pushed the boundaries of what agricultural robots can perceive in the field.
Methodology
Overall Architecture
The researchers built their detector on an Efficient Layer Aggregation Network (ELAN) backbone, introducing two core architectural innovations to handle the specificities of orchard imagery. The model is structured into three main stages: the Backbone for feature extraction, the Feature Pyramid for enhancement and fusion, and the Classification and Regression head for final detection.
Multi-scale Texture Enhancement Module (TEM)
Placed between the backbone and the feature pyramid, the TEM is designed to capture fine-grained details that color-based features might miss. When targets are camouflaged (e.g., green fruit on green leaves), texture becomes the primary tie-breaker. The TEM uses four parallel branches with different dilated convolution rates, allowing the network to perceive textures at multiple scales simultaneously. This helps distinguish the specific surface patterns of a fruit or flower from the surrounding foliage.
Adaptive Region-Aware Feature Fusion
Instead of simple feature stacking, the model employs a dual-attention strategy to refine information flow:
- Selective Region Attention Module (SRAM): Uses high-level semantic information to guide low-level spatial features, effectively providing "top-down" guidance on where the model should focus.
- Selective Channel Attention Module (SCAM): Evaluates the importance of different feature channels to prioritize relevant information, such as edge or texture data, for the specific detection task.
Experimental Evaluation
The model was rigorously tested on four distinct datasets to prove its robustness and generalization:
- Kiwi Flower Dataset: 900 high-resolution images of flowers in various states (unopened, pollinated, etc.).
- Kiwi Fruit Dataset: 2,508 images categorized by "pickability" (single, clustered, or occluded).
- VisDrone2019: A standard benchmark for small object detection from aerial footage.
- MS COCO: The gold standard for general-purpose object detection.
Key Results
The quantitative improvements were substantial. In kiwi flower detection, the model achieved an [email protected] of 0.991, a 2.4% improvement over the YOLOv7 baseline. On the more challenging kiwi fruit dataset, it reached an [email protected]:.95 of 0.709, outperforming the larger YOLOv7x (0.672). Notably, the model achieved these results with significantly fewer parameters (44.7M vs. 70.8M for YOLOv7x), making it more suitable for deployment on robotic hardware.
Conclusion
This research demonstrates that specialized modules for texture enhancement and adaptive fusion can significantly improve detection performance in complex, camouflaged environments. While the current work focuses on the vision pipeline, future efforts will likely involve integrating these high-accuracy detections with robotic control systems for automated picking and pollination. Further optimization, such as model quantization, may also be explored to ensure real-time performance on low-power edge devices in the field.
Summary updated