💻 computer science

Lightweight Adaptive YOLOv11 for Industrial Defect Perception and Closed-Loop Sorting Cybernetic Control System

This paper proposes a lightweight adaptive YOLOv11 detector integrated with a closed-loop cybernetic control system to achieve high-speed, high-accuracy industrial defect perception and real-time sorting, significantly improving inference efficiency and control stability compared to conventional methods.

Original authors: shiqi wang

Published 2026-07-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: shiqi wang

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 factory assembly line where steel sheets are moving fast. The goal is to spot tiny scratches, dents, or spots on the metal and automatically push the bad ones off the line while letting the good ones pass. This is the job of the system described in this paper.

Here is the story of how the author, Shiqi Wang, built a smarter, faster, and more connected way to do this.

The Problem: The Old Way Was Too Clunky and Slow

Traditionally, factories either used human workers to look at the metal (who get tired and miss things) or used old-school computer programs that were like "rigid robots." These robots followed strict rules and couldn't handle tricky lighting or weirdly shaped defects.

Then, people started using YOLO (a famous AI that spots objects in pictures). The latest version, YOLOv11, is very good at spotting things. However, the author found three big problems with using it straight out of the box for a factory:

It's too heavy: The AI brain is too big for the small, cheap computers (called "edge devices") that sit right on the factory floor.
It gets confused: If a defect is tiny and blurry, or if the metal is moving fast, the AI sometimes misses it or gets distracted by shadows.
It talks to no one: Most AI studies just say, "I see a scratch!" and then stop. They don't actually do anything about it. They are like a security guard who sees a thief but doesn't call the police.

The Solution: A "Lightweight" and "Adaptive" AI

The author created a new version of YOLOv11 that is like a super-fit athlete instead of a heavy weightlifter. He made two main upgrades:

1. The "Smart Glasses" (Lightweight Coordinate Adaptive Attention)
Imagine trying to find a tiny speck of dust on a shiny car in the dark. A normal camera might get confused by the glare.
The author added a special module called LCAA. Think of this as a pair of "smart glasses" for the AI. Instead of just looking at the whole picture, these glasses focus specifically on the coordinates (the exact X and Y location) of the tiny defects.

The Result: The AI gets much better at spotting tiny, blurry scratches without needing a massive computer to do the math. It's like giving the AI a magnifying glass that only weighs a feather.

2. The "Chameleon Neck" (Adaptive Multi-Scale Feature Fusion)
Factory defects come in all sizes: some are huge scratches, others are tiny pinpricks. The old AI used the same "recipe" to look at both, which didn't work well.
The author gave the AI a chameleon neck. This part of the system can change its strategy on the fly.

If it sees a tiny defect, it zooms in on the fine details (like looking at the texture of a leaf).
If it sees a huge defect, it zooms out to see the big picture (like looking at the shape of a mountain).
The Result: The AI stops missing small things and stops getting confused by big things. It adapts to whatever it sees in real-time.

The Big Leap: Connecting the Brain to the Hand

This is the most important part of the paper. The author didn't just stop at making a better AI detector. He built a closed-loop system.

Think of it like a reflex arc in your body:

Eyes (Perception): The AI sees a defect.
Brain (Decision): The computer instantly decides, "That's bad, push it off!"
Hand (Execution): A mechanical arm physically pushes the bad steel off the conveyor belt.

In this system, the AI isn't just a camera; it's the sensor for a robot. The paper claims this whole process happens in less than 50 milliseconds (that's faster than a human blink). The AI sees the defect and tells the robot to move before the steel even leaves the camera's view.

The Results: How Well Did It Work?

The author tested this system on a public dataset of steel defects (the NEU dataset) and on a physical test platform.

Accuracy: It found defects 97.2% of the time (very high).
Speed: It was 32% faster than the original YOLOv11.
Size: It was 28% smaller (lighter), meaning it fits easily on small factory computers.
Reliability: In the physical test, the system successfully sorted 98.5% of the items with almost no errors, keeping the production line running smoothly.

The Bottom Line

This paper isn't just about making a better "object detector." It's about connecting the eyes of the AI to the hands of the robot.

The author created a system that is light enough to run on small computers, smart enough to spot tiny defects in messy factory conditions, and fast enough to physically sort bad products off a moving line in real-time. It turns a simple image-recognition tool into a full, self-correcting industrial control system.

Technical Summary: Lightweight Adaptive YOLOv11 for Industrial Defect Perception and Closed-Loop Sorting Cybernetic Control System

Problem Statement
Industrial surface defect inspection and automatic sorting are critical feedback links in intelligent manufacturing cybernetic systems. However, existing detection methods face three primary bottlenecks when applied to industrial cybernetic control:

Deployment Constraints: Conventional models often possess bulky parameter counts and low inference speeds, failing to meet the millisecond-level feedback constraints required for real-time industrial control on low-power edge devices.
Environmental Robustness: Industrial defects exhibit multi-scale distributions and weak feature signals susceptible to uneven illumination, surface occlusion, and background noise. Fixed-structure networks often suffer from unstable feature extraction, leading to false or missed detections.
System Disconnection: Most existing YOLO-based research focuses solely on isolated image recognition tasks. There is a lack of coupling between detection outputs and industrial actuators, preventing the formation of a standardized perception-decision-execution closed-loop cybernetic feedback system.

Methodology
The paper proposes a Lightweight Adaptive YOLOv11 detector integrated into a complete machine-learning-perception-driven closed-loop sorting cybernetic control system. The approach involves three core technical components:

Lightweight Coordinate Adaptive Attention (LCAA) Module:
- Embedded into the YOLOv11 backbone's C2f blocks.
- Integrates spatial coordinate perception with channel adaptive weighting.
- Utilizes horizontal and vertical one-dimensional pooling to capture long-distance spatial dependencies, avoiding the computational overhead of 2D global pooling.
- Replaces traditional fully connected layers with a lightweight $1\times1$ convolution to reduce parameters by approximately 15% while enhancing the feature response of tiny, blurred defects and suppressing background noise.
Adaptive Multi-Scale Feature Fusion Strategy:
- Implemented in the neck network (Feature Pyramid).
- Dynamically adjusts fusion weights between shallow fine-grained textures and deep semantic features based on the real-time scale distribution of target defects within a batch.
- Uses a scale factor $s$ derived from bounding box dimensions to calculate adaptive weights ( $\alpha_s$ and $\alpha_d$ ). This ensures tiny defects retain high-resolution texture information, while large defects leverage global semantic integrity, mitigating missed detections across varying scales.
Closed-Loop Cybernetic Control System:
- Architecture: Comprises an edge visual perception module (Jetson Nano), an upper computer decision control module, and an execution sorting module (mechanical arm).
- Logic: The optimized model acts as the system sensor. Real-time detection outputs (defect type, coordinates, confidence) serve as state feedback signals.
- Control: The upper computer matches detection data against preset logic to generate sorting instructions. The mechanical arm executes offset sorting for defective products while allowing qualified products to pass, forming a negative feedback loop.
- Deployment: The model is quantized and optimized for edge embedded hardware to ensure single-frame detection times under 30ms, with a total system end-to-end delay under 50ms.

Key Contributions
The paper identifies three core interdisciplinary innovations:

Algorithmic Optimization: The proposal of the LCAA module, which simultaneously compresses network computational overhead and enhances feature extraction for tiny, low-quality defects, ensuring stable input signals for the controller.
Dynamic Adaptation: The design of an adaptive multi-scale feature fusion strategy that autonomously adjusts feature fusion proportions based on target scale, improving robustness against variable defect sizes and preventing control signal jitter caused by detection instability.
System Integration: The construction of a complete "perception-decision-execution" closed-loop framework. This work docks real-time lightweight detection outputs directly with industrial sorting controllers, breaking the separation between machine learning algorithms and industrial automatic control equipment.

Experimental Results
Experiments were conducted on the public NEU steel defect dataset (1,800 images, 6 defect types) and validated on a physical edge sorting platform.

Detection Performance: The proposed method achieved 97.2% mAP@0.5, an improvement of 2.7% over vanilla YOLOv11 (94.5%).
Efficiency: Compared to the original YOLOv11, parameters were reduced by 28.6% (from 7.82M to 5.58M), and inference speed (FPS) increased by 32.1% (from 58.6 to 77.4).
Comparative Analysis: The method outperformed mainstream algorithms including SSD, YOLOv8, and YOLOv10 in terms of accuracy, parameter count, and inference speed.
Control System Performance: Physical platform tests demonstrated a sorting success rate of 98.5% and a steady-state error of less than 2%. The system maintained an overall end-to-end delay below 50ms, satisfying the stability constraints of discrete industrial production line control.

Significance and Claims
The paper positions this work as a solution to the disconnection between machine learning visual detection and industrial automatic control.

Interdisciplinary Integration: It claims to realize the organic integration of machine learning perception and industrial cybernetic closed-loop regulation, aligning with the interdisciplinary positioning of machine learning and cybernetics.
Engineering Value: By treating the detection model as a sensing unit within a discrete control system, the study moves beyond pure computer vision to provide a standardized paradigm for visual perception feedback control in intelligent manufacturing.
Practical Application: The system enables online quality statistics and dynamic production line speed adjustment, offering a feasible technical scheme for the integration of visual perception and cybernetic control in various industrial scenarios.

The authors acknowledge limitations, noting that current validation is restricted to the NEU dataset and a simulation platform, and that future work will focus on expanding defect datasets and adding intelligent scheduling and fault self-diagnosis functions.