Lightweight Dual-Path Network with Dynamic Distillation for Industrial LED Defect Recognition

In the factories that power our modern world, tiny lights called LEDs are assembled by the millions. These components are the eyes of our screens and the beacons of our cities, but they are fragile. During manufacturing, they can develop subtle flaws: a bubble trapped in the glue, a pin that is slightly loose, or a speck of foreign dust. If these defects slip through, the final product may fail, leading to costly waste and unreliable technology. For decades, manufacturers have tried to catch these errors using automated cameras and computers. Early attempts relied on rigid rules or simple pattern matching, which often failed when lighting changed or the defects looked slightly different. More recent systems use deep learning, a type of artificial intelligence that learns to recognize patterns by studying thousands of examples. However, these smart systems often require massive computing power, making them too heavy and slow to run on the small, affordable computers found on a busy factory floor. Furthermore, real-world factory data is messy; some types of defects are common, while others are rare, and different defects can look startlingly similar, confusing even the most advanced algorithms.

To solve this, a team of researchers from several universities in China has developed a new, lightweight system designed specifically for these difficult industrial conditions. Their approach, detailed in a recent study, focuses on teaching a small, efficient computer model to recognize LED defects with high precision, even when the data is unbalanced and the visual clues are faint. Instead of relying on a single, rigid way of looking at an image, the new system uses a "dual-path" strategy. Imagine a student learning a complex subject: they might study a textbook for the big picture while also examining detailed diagrams for the fine points. This system does something similar by running two parallel streams of analysis. One stream looks at the overall structure of the image, while the other zooms in on specific layers of detail. Crucially, the system does not treat these two views as equally important for every single image. Instead, it uses a dynamic mechanism to decide, for each specific defect, how much weight to give to the broad view versus the detailed view. This allows the computer to adapt its focus, paying extra attention to the details when two different defects look very similar, or relying on the broader context when the defect is obvious.

To make this small system as smart as possible, the researchers employed a technique called knowledge distillation. In this process, a large, powerful, and computationally expensive "teacher" model is first trained to recognize the defects perfectly. This teacher then guides the smaller "student" model, which is the one intended for the factory. The teacher does not just tell the student the correct answer; it shares the subtle relationships between different types of defects and the specific features that distinguish them. The student learns from both the final decisions of the teacher and the intermediate steps the teacher took to reach those decisions. This transfer of knowledge allows the lightweight student to achieve a level of accuracy that usually requires a much larger machine. The researchers tested this method on two different sets of data: a new dataset they built containing over 6,000 images of eight different defect types, and an existing public dataset used by a major technology company. The results were striking. On their own dataset, the system correctly identified defects 97.17% of the time, and on the public dataset, it reached 96.31% accuracy. These numbers represent a significant improvement over other lightweight models currently available, which often struggle with the rare and confusing defect types found in real production lines.

The study also addressed the challenge of limited computing resources, a critical factor for industrial deployment. While the new system is slightly more complex than the simplest existing models, it remains small enough to run efficiently on standard hardware. It processes images at a speed of over 300 frames per second, fast enough to keep up with the rapid pace of a manufacturing line. The researchers demonstrated that the system does not just memorize the training images but learns to generalize, meaning it can identify defects it has never seen before with high reliability. By combining the dual-path architecture with the knowledge distillation technique, the team created a tool that is both powerful and practical. This work suggests that high-precision quality control does not require massive, energy-hungry supercomputers. Instead, by teaching smaller models to learn more effectively from their larger counterparts and to adapt their focus dynamically, factories can achieve better defect detection without sacrificing speed or efficiency. The researchers have made their code and the new dataset available to the public, inviting others to build upon this foundation for smarter, more reliable industrial manufacturing.