💻 computer science

The Lightweight Cherry Tomato Fruit Maturity Detection Method Based on Improved YOLOv10n

This paper proposes YOLOv10n-FBD, a lightweight cherry tomato maturity detection model that integrates CCFM feature fusion, PSA-BiFormer attention, and a C2f-Dual strategy to achieve high precision (94.3% mAP) and speed (369 FPS) while significantly reducing model size and computational complexity compared to existing methods.

Original authors: Ji Cai, Shuaishuai Cui, Baofan Chen, Yuhao Hao, Kun Wang, Guozhu Song

Published 2026-06-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ji Cai, Shuaishuai Cui, Baofan Chen, Yuhao Hao, Kun Wang, Guozhu Song

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

Imagine a busy cherry tomato farm. The farmers have a tough job: they need to pick the tomatoes at the exact right moment. If they pick them too early, they are sour; too late, they rot. Doing this by hand is slow, expensive, and often hurts the fruit.

This paper introduces a new "smart camera brain" designed to help robots or farmers instantly know if a cherry tomato is Green (unripe), Red (ripe), or Reddish (semi-ripe). The goal was to build a system that is super accurate, incredibly fast, and small enough to fit on a simple device, like a smartphone or a small robot, without needing a massive supercomputer.

Here is how the researchers built this "smart brain," explained through simple analogies:

The Starting Point: The "Base Model"

The researchers started with a standard, high-tech detection system called YOLOv10n. Think of this as a very talented, but slightly heavy, detective. It's good at finding things, but it's a bit slow and takes up a lot of memory space on a device. The team wanted to make this detective faster and lighter without losing its sharp eyes.

The Three Upgrades (The "Secret Sauce")

To create their new model, called YOLOv10n-FBD, they gave the detective three specific upgrades:

1. The "Multi-View Binoculars" (CCFM Feature Fusion)

The Problem: Sometimes tomatoes are tiny in the distance, or hidden behind leaves. The original model might miss the small details or get confused by the background.
The Fix: They added a CCFM module. Imagine giving the detective a set of binoculars that can look at the scene through different lenses simultaneously—one lens for close-up details (like the texture of the skin) and another for the big picture (like the overall color of the cluster).
The Result: By mixing these different views, the model gets a much richer understanding of the fruit, making it much better at spotting small or partially hidden tomatoes.

2. The "Spotlight" (PSA-BiFormer Attention)

The Problem: In a field, there is a lot of noise: green leaves, brown stems, and bright sunlight. The original model sometimes wastes energy looking at the leaves instead of the fruit.
The Fix: They replaced the old attention system with a BiFormer mechanism. Think of this as a smart spotlight. Instead of scanning the whole field evenly, the spotlight instantly locks onto the red or reddish shapes (the tomatoes) and ignores the green leaves and stems. It's like a bouncer at a club who only lets the VIPs (the tomatoes) through and ignores everyone else.
The Result: The model focuses its energy exactly where it needs to, improving accuracy significantly.

3. The "Lightweight Backpack" (C2f-Dual Strategy)

The Problem: The original model was heavy. It had too many "muscles" (parameters), making it slow and hard to run on small devices.
The Fix: They used a C2f-Dual strategy. Imagine the detective was carrying a heavy backpack full of tools they never used. The researchers swapped the heavy tools for a sleek, dual-purpose Swiss Army knife. They split the work into two efficient paths: one path handles the shape (contours), and the other handles the color.
The Result: They cut the model's size by nearly 40% and made it 37.5% lighter in terms of data, but it didn't lose any of its detective skills.

The Results: A Super-Performing Detective

After training this new system on thousands of photos of cherry tomatoes, the results were impressive:

Speed: It can process 369 images per second. To put that in perspective, it's like watching a movie at 369 frames per second—blazing fast. It can count and classify tomatoes in real-time as a robot moves through a field.
Accuracy: It correctly identified the maturity of tomatoes 94.3% of the time (using a strict standard). This is a big jump compared to older models.
Size: The entire "brain" is only 3.5 MB. That's smaller than a single high-quality photo you might take with your phone, meaning it can easily fit on small, cheap hardware.

The Limitations (What the Paper Admits)

The authors are honest about where their system might struggle. Because they wanted to make sure the training data was perfect, they excluded tomatoes that were:

Hidden deep behind leaves (heavily occluded).
Very far away or tiny.
In extreme lighting (like blinding glare or pitch black).

They admit that if a robot uses this in a messy, real-world orchard with lots of leaves and weird angles, it might make a few more mistakes than it did in their controlled tests. They also noted that combining all three upgrades made the model slightly less accurate at spotting the "semi-ripe" tomatoes compared to using just two of the upgrades, suggesting there is still a tiny bit of room for fine-tuning.

Summary

In short, the researchers took a powerful but heavy AI model and gave it better glasses (to see details), a smart spotlight (to ignore distractions), and a lightweight backpack (to run fast). The result is a system that can instantly tell a farmer if a cherry tomato is ready to be picked, doing it faster and more efficiently than previous methods, all while fitting on a small device.

Technical Summary: The Lightweight Cherry Tomato Fruit Maturity Detection Method Based on Improved YOLOv10n

Problem Statement

The detection and recognition of cherry tomato fruit maturity face a critical challenge: achieving a simultaneous balance between high precision, fast recognition speed, and a small model size. Traditional harvesting methods are labor-intensive and inefficient, while existing intelligent harvesting systems often suffer from insufficient detection accuracy, high miss rates, and excessive computational complexity. Large model sizes and high processing demands hinder the deployment of these systems on resource-constrained devices, limiting their practical application in real-world agricultural environments. While various deep learning approaches (e.g., MobileNetV3-based YOLOv7, DSP-YOLOv7-CA, Swin Transformer optimizations) have shown promise, many struggle to find an optimal equilibrium among speed, accuracy, and parameter count.

Methodology

This study proposes a lightweight detection method, YOLOv10n-FBD, built upon the baseline YOLOv10n architecture. The model addresses the identified challenges through three primary technical improvements:

CCFM (Channel Cross Feature Modulation) Feature Fusion:
To mitigate information loss in small-scale object detection caused by limited pixels and repeated convolution/pooling, the authors replaced the original feature fusion component with the CCFM module (adapted from RT-DETR). This module utilizes a bidirectional feature flow architecture similar to PANet but incorporates Fusion Blocks. These blocks employ RepBlocks and attention mechanisms to fuse multi-scale feature maps ( $S3, S4, S5$ ) from the backbone. By flattening high-level features and applying multi-head self-attention to recalibrate channel weights, CCFM enhances the integration of fine-grained local details with global contextual information, improving adaptability to scale variations.
PSA-BiFormer Attention Mechanism:
To overcome the high computational complexity and memory demands of the original Position-Sensitive Attention (PSA) mechanism, the study integrates the BiFormer attention mechanism. BiFormer employs a Bi-level Routing Attention (BRA) strategy. First, a region-level affinity graph is constructed and pruned to retain only the top- $k$ connections for each node, filtering out irrelevant regions. Second, a token-to-token attention mechanism is applied using dense matrix multiplication on the filtered key-value pairs. This approach achieves content-aware sparsity, allowing the model to focus on critical features (e.g., red fruits) while suppressing background interference (e.g., green leaves) without the heavy computational cost of global attention.
C2f-Dual Lightweight Strategy:
To reduce computational complexity and model size, the standard C2f module in the backbone was restructured using a DualConv design. This strategy splits the convolution operation into two parallel sublayers:
- A 3×3 convolutional branch to capture spatial features and texture details.
- A 1×1 convolutional branch to integrate channel interactions and semantic information.
  These branches utilize group convolution and fusion strategies to significantly reduce parameters and FLOPs while maintaining gradient flow integrity through residual connections.

Key Contributions

Model Architecture: The construction of the YOLOv10n-FBD model, which successfully integrates CCFM, BiFormer, and C2f-Dual strategies to optimize the trade-off between accuracy and efficiency.
Feature Enhancement: The introduction of CCFM to improve multi-scale feature representation, specifically aiding in the detection of partially occluded or small cherry tomatoes.
Efficient Attention: The adaptation of the BiFormer mechanism to replace standard PSA, achieving efficient global feature processing through a two-level routing strategy that reduces redundant calculations.
Lightweight Optimization: The implementation of the C2f-Dual strategy, which significantly compresses the model size and parameter count without compromising detection performance.

Experimental Results

The model was trained and evaluated on a dataset of 4,046 cherry tomato images collected from Shanxi Agricultural Valley, categorized into three maturity stages: Ripe (R), Unripe (G), and Semi-ripe (RG).

Performance Metrics: The YOLOv10n-FBD model achieved an mAP@.5:.95 of 94.3%, representing a 2.2% improvement over the traditional YOLOv10n baseline.
Efficiency Gains:
- Parameter Count: Reduced by 37.5%.
- Model Size: Compressed to 3.5 MB, a decrease of 38.6%.
- Computational Cost: Reduced to 6.0 GFLOPs.
- Inference Speed: Achieved 369.0 FPS on an NVIDIA RTX 3090 GPU.
Comparative Analysis: Compared to mainstream models (Faster R-CNN, SSD, YOLOv5, YOLOv7, YOLOv8, and YOLOv11), the proposed model demonstrated superior performance in mAP@.5:.95, with improvements ranging from 2% to 15%. It also showed significant reductions in parameters and model size compared to all compared YOLO variants.
Ablation Studies: Experiments confirmed that while the combination of all three modules (CCFM + BiFormer + C2f-Dual) resulted in a slight decrease in mAP@.5 (97.9%) compared to the CCFM+BiFormer combination (98.4%), the stricter mAP@.5:.95 metric remained stable at 94.3%. This suggests the combined model offers better robustness and localization accuracy under rigorous evaluation criteria.

Significance and Limitations

The paper claims that the YOLOv10n-FBD model successfully constructs an accurate and efficient recognition network that balances precision and efficiency, making it suitable for complex, variable real-world application environments and resource-constrained devices. The model demonstrates high robustness in dynamic video validation scenarios, maintaining high confidence scores and detection rates.

However, the authors modestly acknowledge specific limitations:

Dataset Constraints: The dataset excluded occluded, distant, and extremely small fruits to ensure annotation consistency. Consequently, the dataset may not fully reflect the complexity of dense orchard environments, potentially leading to an underestimation of detection errors in such scenarios.
Generalization: Data collection was confined to specific lighting conditions, angles, and a single geographical location (Shanxi Agricultural Valley), which may limit the model's generalization to extreme lighting or different planting patterns.
Module Compatibility: The ablation studies revealed a minor trade-off where the lightweight C2f-Dual design slightly impacted the precision of the BiFormer attention mechanism when all modules were combined, indicating room for further optimization in module compatibility.

The study concludes that while the proposed method significantly advances lightweight maturity detection, future work will focus on relaxing annotation rules to include occluded targets, augmenting datasets for diverse environments, and further refining the synergy between the introduced modules.