A Lightweight Lychee Detection Network for Orchard Harvesting Robots via ODConv and Spatial Feature Restoration
This paper proposes ODAF-YOLOv11-n, an ultra-lightweight object detector that integrates Omni-Dimensional Dynamic Convolution and a Lightweight Spatial Feature Restoration Module to achieve high-accuracy, real-time lychee detection in complex orchard environments while significantly reducing computational complexity for edge robotic deployment.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the lush, humid orchards of southern China, the lychee fruit hangs in heavy, fragrant clusters, a prized harvest that has long relied on human hands to gather. Yet, as rural labor becomes scarcer and the demand for efficiency grows, farmers are turning to robots to do the picking. For a machine to harvest a fruit, it must first see it clearly. This is the domain of computer vision, a field where cameras and software work together to identify objects in the world. The challenge in an orchard is that nature is rarely cooperative; the light shifts from blinding sun to deep shadow, and the fruit is often hidden behind layers of leaves and other branches. To succeed, a robot needs a "brain" that can process these images instantly, even when the view is cluttered or the lighting is poor, all while running on a small computer attached to the robot itself.
Researchers have been developing these visual systems for years, often using complex mathematical models known as deep learning networks. These systems learn to recognize patterns by studying thousands of images, but they face a difficult trade-off. The most accurate systems are often too heavy and slow for a robot to carry, while the light, fast systems often fail when the environment gets messy. A team of researchers from South China Business College and Guangdong Polytechnic of Industry and Commerce has now proposed a new solution designed specifically for this problem. They created a streamlined visual system called ODAF-YOLOv11-n, which is built to find lychees in the chaotic, unstructured environment of a real orchard without needing a supercomputer to run it.
The core of their work addresses three specific hurdles that have plagued previous attempts. First, the orchard light is unpredictable. A robot might be working in the bright glare of midday sun, where fruit skins reflect harsh highlights, or in the dim, green shadows of the canopy. Standard computer vision tools often struggle here, losing the ability to distinguish the fruit from the leaves. To fix this, the researchers embedded a special mechanism into the robot's visual processing that acts like a dynamic filter. Instead of using a fixed set of rules to look at an image, this system adjusts its own sensitivity in real-time. It can shift its focus to amplify faint textures in the dark or suppress blinding glare in the bright, ensuring the robot sees the fruit clearly regardless of the weather.
The second challenge is physical obstruction. Lychees grow in tight bunches, frequently buried under leaves or pressed against other fruits. When a fruit is only partially visible, standard systems often miss it entirely because they cannot guess the shape of the hidden part. The researchers tackled this by adding a "restoration" step to their system. Imagine looking at a puzzle where several pieces are missing; a standard system might give up, but this new system uses the visible edges and the surrounding context to mentally reconstruct the missing parts. It analyzes the relationship between different layers of the image, effectively filling in the gaps to locate fruits that are heavily covered by foliage. This allows the robot to identify a lychee even when only a small fraction of it is visible.
Finally, the team had to ensure the system was light enough to run on a robot's onboard computer. Previous high-accuracy models were too bulky, requiring massive amounts of energy and processing power that would drain a robot's battery or slow down its movements. The researchers solved this by stripping away unnecessary computational weight from the system. They replaced heavy, standard processing steps with a much more efficient method that handles spatial and channel information separately, drastically reducing the number of calculations needed. This allowed them to keep the system small and fast while still retaining the powerful features needed to handle the difficult lighting and occlusion issues.
The results of this work were tested on two large collections of lychee images, one from a public dataset and another captured specifically in the Conghua district of Guangzhou. The new system proved to be remarkably effective. On the public dataset, it achieved a detection accuracy of 95.53%, and on the more difficult, real-world dataset, it reached 88.13%. Perhaps most importantly, it did this while using only 2.96 billion calculations per second and occupying just 2.94 million parameters of memory. This is a significant reduction in complexity compared to the standard models it was based on, which required more than double the computational power to achieve lower accuracy.
The system's true strength was revealed when the researchers tested it under extreme conditions. When they simulated an environment where 80% of the fruit was hidden by leaves, the new system still managed to find the targets with a success rate of 32.75%, far outperforming other leading models that dropped below 22%. Similarly, in tests involving extreme brightness and deep shadows, the system maintained high precision and recall, proving that its adaptive mechanisms work as intended. The researchers found that by combining these three improvements—dynamic light adaptation, spatial feature restoration, and computational efficiency—they created a tool that bridges the gap between complex artificial intelligence and the practical needs of field robotics.
This work suggests that the future of automated harvesting may not require massive, slow computers, but rather clever, lightweight systems designed to understand the messy reality of nature. The researchers acknowledge that their current work focuses on two-dimensional vision and that future steps will involve integrating depth sensors to help robots grasp the fruit in three-dimensional space. They also plan to address the issue of moving targets, as wind and robot movement can cause the fruit to sway. However, the current findings demonstrate a significant step forward, offering a reliable visual foundation that could allow robots to harvest lychees efficiently in the unstructured, challenging environments where they naturally grow.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.