💻 computer science

A Lightweight YOLOv11n-OBB-Based Method for Oriented Detection of Luosifen Outer Packages

This paper proposes a lightweight YOLOv11n-OBB-based detection method featuring a Surgical Precision replacement strategy with GhostConv_OBB and C3k2_GhostConv_OBB modules, which achieves high-precision, real-time oriented detection of Luosifen outer packages on edge devices while significantly outperforming the baseline in accuracy and efficiency.

Original authors: Guojian Liang, Shuwen Zhou, Yinhai Li

Published 2026-07-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Guojian Liang, Shuwen Zhou, Yinhai Li

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 factory assembly line where robots are tasked with grabbing bags of Luosifen (a famous spicy river-snail rice noodle) and packing them into boxes. To do this successfully, the robot's "eyes" (a camera system) need to do three things at once:

Spot the bag among many different types.
Figure out exactly how it's tilted (is it lying flat, standing up, or spinning?).
Do it instantly on a small, low-power computer attached to the robot, not a massive supercomputer.

This paper presents a new "eye" for these robots that is incredibly fast, small, and accurate. Here is how they did it, explained simply:

The Problem: The "Box" vs. The "Bag"

Standard computer vision usually draws a square box around objects. Imagine trying to grab a bag of chips that is lying diagonally on a table. If you draw a square around it, you also grab a lot of empty table space. The robot gets confused, grabs the wrong spot, or misses the bag entirely.

To fix this, engineers use Oriented Bounding Boxes (OBB). Instead of a square, the box rotates to hug the exact shape of the bag. However, most existing "rotating box" systems are like heavy, fuel-guzzling trucks. They are too big and slow to run on the small, cheap computers (edge devices) that sit right on the factory floor.

The Solution: A "Ghost" Trick with a Safety Net

The researchers built a new system based on a very lightweight model called YOLOv11n-OBB, but they gave it a major upgrade using three clever tricks:

1. The "Ghost" Chef (GhostConv_OBB)

Imagine a chef who needs to prepare a huge meal but only has a tiny kitchen. Instead of hiring 100 cooks, they hire one main chef and a few "ghost" assistants who just copy the main chef's work with very little effort.

How it works: The system uses a technique called GhostConv. It generates most of the visual data using a cheap, low-effort method, creating "ghost" features that look real but cost almost nothing to compute. This keeps the system tiny and fast.

2. The "Flashlight" and the "Noise-Canceling Headphones" (AAE & PECA)

Even with a small system, the factory floor is messy. Bags are crinkled, shiny, and sometimes stacked on top of each other.

The Flashlight (AAE): This module acts like a focused flashlight. It specifically shines light on the edges and corners of the noodle bags. This helps the robot understand the bag's shape and rotation angle, even if the bag is twisted or wrinkled.
The Noise-Canceling Headphones (PECA): The factory is loud with visual "noise" (shiny reflections, conveyor belt patterns). This module acts like headphones that cancel out that background noise, letting the robot focus only on the bag itself.
The Magic: The paper found that using the flashlight and the headphones together works much better than using just one. The flashlight finds the bag, and the headphones ensure the robot doesn't get distracted by the background.

3. The "Goldilocks" Placement (Key-Layer Sensitivity)

The researchers didn't just throw these upgrades everywhere. They tested 71 different places to put the new modules to see where they worked best.

The Result: They found the "Goldilocks" spot: the Neck-to-Head transition layer (specifically detection head 11).
The Analogy: Imagine a relay race. If you give the baton to a runner too early, they might trip on the starting line. If you give it too late, they might miss the finish. They found the perfect moment in the middle of the race to swap in their new, faster runner, ensuring the robot gets the best possible view right before making the decision to grab.

4. The "Safety Net" (Numerical Stability)

Running software on small, cheap computers can sometimes lead to "math errors" (like a calculator dividing by zero).

The Fix: They added a "safe fallback" mechanism. If the math gets weird or breaks, the system instantly switches to a simple, safe backup mode (like a parachute) so the robot doesn't crash or stop working. This ensures the factory line keeps running 24/7.

The Results: A Small Engine with a Ferrari's Speed

The team tested their new system against the original YOLOv11 and other heavy-duty models.

Accuracy: It got 93.95% accuracy on the strictest test (meaning it grabbed the bag perfectly, not just "close enough"). This is a huge jump of over 32 percentage points compared to the standard version.
Size: It is still tiny, using only 2.95 million parameters (about the size of a small smartphone app).
Speed: It runs at 14.76 frames per second, which is fast enough for the robot to grab bags in real-time without stuttering.

Summary

The paper describes a way to make a robot's vision system small enough to fit on a cheap chip but smart enough to handle messy, spinning noodle bags. By using "ghost" features to save space, a "flashlight" to find edges, "noise-canceling" to ignore distractions, and placing these tools in the perfect spot, they created a system that is both lightweight and incredibly precise. This allows factories to automate the boxing of Luosifen without needing expensive, heavy computers.

Technical Summary: A Lightweight YOLOv11n-OBB-Based Method for Oriented Detection of Luosifen Outer Packages

Problem Statement
The automated boxing of Luosifen (river-snail rice noodle) outer packages on production lines requires a detection system capable of simultaneous multi-category detection, rotation-angle estimation, and real-time inference under strict edge-computing constraints. Existing solutions face three primary limitations:

Template Matching: Conventional methods struggle with flexible production scenarios involving multiple object poses and cannot directly output rotation angles.
Horizontal Bounding Boxes (HBB): While efficient, HBBs introduce significant background interference for tilted objects and fail to estimate rotation angles, leading to inaccurate robotic grasping points.
Standard Oriented Bounding Box (OBB) Models: Although OBBs address rotation, standard models often suffer from excessive backbone redundancy and high computational overhead, making them difficult to deploy on resource-constrained edge devices. Furthermore, existing lightweight OBB methods often lack sufficient feature extraction for rotated edges, struggle with complex industrial backgrounds, and lack fault-tolerant mechanisms for numerical stability during long-term edge inference.

Methodology
The authors propose a lightweight oriented detection method based on a key-layer sensitivity-guided replacement strategy, using YOLOv11n-OBB as the baseline. The approach focuses on three core objectives: precise lightweight optimization, angle-aware enhancement, and progressive background suppression.

Dataset Construction:
- A specialized dataset of 1,603 images covering four categories of Luosifen outer packages was created.
- Data was collected in both static and dynamic (conveyor-belt) states to simulate motion blur and vibration.
- The dataset was augmented (rotation, perspective, illumination, occlusion) and split 7:2:1 (1,122 training, 320 validation, 161 test images).
- Annotations use parametric oriented bounding boxes (OBBs) to cover arbitrary in-plane orientations.
Architectural Improvements:
- GhostConv_OBB Unit: Replaces standard convolutions in the backbone. It utilizes a cost-efficient GhostConv mechanism to generate intrinsic and supplementary "ghost" features. Crucially, it integrates:
  - Dual-Path Angle-Aware Enhancement (AAE): A global-local dual-path module that strengthens feature responses along package edges and corners to provide geometric priors for angle estimation.
  - Progressive Efficient Channel Attention (PECA): A two-stage convolutional structure (kernel sizes 5 and 3) that progressively learns channel weights to suppress background noise without fully connected layers.
  - Learnable Residual Fusion: Replaces fixed identity mapping with a learnable coefficient ( $\lambda$ ) to adaptively balance residual and enhanced features, including a safe fallback mechanism for numerical anomalies (e.g., NaN/Inf).
- C3k2_GhostConv_OBB Module: Replaces the standard C3k2 module in the Neck to reduce parameter count during multi-scale feature fusion while preserving cross-layer shortcuts.
- GhostBottleneck_OBB: A deep residual propagation unit stacking GhostConv_OBB subunits with learnable scaling coefficients for stability.
Key-Layer Sensitivity Strategy:
- Instead of global replacement, the authors conducted an ablation study across 71 configurations to identify the optimal deployment position.
- Detection Head 11 (the Neck-to-Head transition layer) was identified as the optimal location. This position offers a balance between receptive field coverage (macro-level geometric information) and gradient flow, avoiding the disruption of low-level edge features or the amplification of numerical errors found in deeper or shallower layers.

Key Contributions

Dataset: Creation of a publicly available, diverse OBB dataset for Luosifen outer packages under dynamic production conditions.
Lightweight Architecture: Proposal of the GhostConv_OBB unit and C3k2_GhostConv_OBB module, which integrate AAE and PECA to enhance rotated object features while maintaining a low parameter count.
Deployment Strategy: Introduction of a key-layer sensitivity-guided replacement strategy, demonstrating that precise modification at Detection Head 11 yields superior accuracy-efficiency trade-offs compared to blind global replacement.
Numerical Stability: Implementation of learnable residual fusion coefficients and a safe fallback mechanism to ensure robust operation on edge devices.

Experimental Results
The proposed model was evaluated against various horizontal and oriented detectors (including YOLOv8, v10, v11, v12, and v26 variants) on the test set:

Performance: The model achieved 97.79% mAP@0.5 and 93.95% mAP@0.5:0.95.
Improvements: Compared to the YOLOv11n-OBB baseline, this represents a 4.26 percentage point increase in mAP@0.5 and a 32.27 percentage point increase in mAP@0.5:0.95.
Efficiency: The model maintains a lightweight footprint with 2.95M parameters and 3.45 GFLOPs.
Latency: Inference latency is 67.73 ms (14.76 FPS), meeting the production line requirement of $\ge$ 14 FPS with minimal overhead compared to the baseline (65.88 ms).
Ablation Insights:
- PECA was found to be the primary driver for strict localization accuracy (improving mAP@0.5:0.95 by 28.20 points over the baseline when used alone).
- AAE primarily improved recall by enhancing edge/corner visibility.
- The combination of AAE and PECA provided complementary benefits, balancing detection recall and localization precision.

Significance and Claims
The paper claims that the proposed method achieves localization accuracy comparable to medium-scale models (e.g., YOLOv11s-OBB) while operating under an extremely low parameter budget suitable for edge deployment. It highlights the complementary synergy between the AAE and PECA modules in improving both recall and localization precision. The work is presented as a lightweight design paradigm for edge-oriented rotated object detection in industrial scenarios, specifically addressing the challenges of complex backgrounds, rotation estimation, and numerical stability on constrained hardware. The authors note that future work will focus on feature decoupling under severe occlusion and INT8 quantization for frameworks like TensorRT.