Joint Enhancement of Multi-Scale Features and Adaptive Loss for YOLOv5
This paper proposes a multi-dimensional collaborative enhancement framework for YOLOv5 that integrates LAB-space CLAHE, a high-resolution P2 detection layer with Varifocal Loss, and deformable convolution with channel attention to effectively address low-light, small-target, and geometric deformation challenges in industrial visual inspection, achieving superior performance over mainstream detectors on a specialized ton bag dataset.
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 world of industrial automation, machines rely on cameras to see what humans see, but they often struggle with the same visual problems that trouble us in dimly lit rooms. When a factory floor is poorly lit, or when a product is crumpled, folded, or tiny, standard computer vision systems can lose their way. They might miss a small tear in a bag or fail to locate a damaged part because the image is too dark or the object is too distorted. This is where the field of object detection comes in, a branch of artificial intelligence designed to teach computers to identify and locate items within an image. While modern systems are powerful, they often falter when faced with the messy reality of a warehouse: low light, tiny defects, and objects that are not perfectly shaped. Researchers are constantly trying to build systems that can handle these difficult conditions without slowing down the fast pace of industrial work.
A team of researchers from Taiyuan University of Science and Technology and Taiyuan Fortucky Logistics Equipment & Technology Co., Ltd. has tackled these specific challenges by improving a popular detection system known as YOLOv5. They focused their work on a particularly tricky task: inspecting industrial ton bags, which are large woven sacks used to transport bulk materials. These bags present a perfect storm of difficulties for a camera. In low-light conditions, the woven texture of the bag becomes hard to distinguish, and any damage, such as a small hole or a tear, is often too small to be seen clearly. Furthermore, as these bags are stacked and moved, they wrinkle and twist, changing their shape in ways that confuse standard cameras. The researchers set out to create a single solution that could handle the darkness, find the tiny flaws, and adapt to the shifting shapes of the bags all at once.
To solve the problem of poor lighting, the team introduced a method that acts like a smart flashlight for the computer's eyes. Instead of simply brightening the entire image, which can wash out details or create blinding spots, they used a technique that adjusts the contrast in small, local areas. Imagine looking at a dark room where some corners are pitch black and others are slightly lit; a simple brightener would make the whole room white and lose the shadows. The method used here, known as CLAHE, looks at tiny patches of the image and boosts the contrast only where it is needed. This allows the computer to see the fine weave of the bag and the subtle differences that indicate a tear, even in very dark corners, while keeping the noise and grain of the image under control. This step ensures that the computer receives a clear, detailed picture before it even begins to look for defects.
Once the image is clear, the system needs to find the tiny damaged areas, which are often just a few pixels wide. Standard detection systems often miss these small targets because they process images in layers that gradually shrink the picture, causing small details to disappear. The researchers added a new layer to the system that keeps a high-resolution view of the image, allowing it to spot these minute flaws without losing them in the blur. They also changed the way the system learns from its mistakes. Instead of treating every error the same, the new method pays extra attention to the hard-to-find small objects and the ones the system is unsure about. This combination of keeping a sharp, close-up view and learning more carefully from difficult examples significantly improved the system's ability to find small tears and holes that it would have previously ignored.
The final challenge was dealing with the bags as they twist and fold. A standard camera lens sees the world in a rigid grid, which works well for straight lines but struggles when an object bends or stretches. To fix this, the researchers gave the system a flexible way of looking at the image. They added a mechanism that allows the computer to shift its focus points slightly, adapting its view to match the wrinkles and curves of the bag. This is paired with a system that learns which parts of the image are most important, effectively telling the computer to ignore the background noise and focus on the shape of the damage. By combining this flexible viewing with a focus on the most relevant features, the system became much better at recognizing damaged bags even when they were crumpled or tilted in strange ways.
When the researchers tested their improved system on a collection of real-world ton bag images, the results were striking. In low-light conditions, the system's ability to detect damage jumped from 65.1% to 78.4%, a significant leap that means far fewer missed defects. For the smallest targets, the detection rate rose to 85.2%, and for deformed bags, it reached 74.3%. Overall, the new system achieved an accuracy of 86.7%, outperforming other leading detection models like YOLOv7 and YOLOv8, which scored 75.8% and 78.6% respectively. The study demonstrates that by addressing lighting, scale, and shape simultaneously, it is possible to create a robust inspection tool that works reliably in the complex and often imperfect environment of an industrial warehouse. This approach offers a practical path forward for ensuring quality control in logistics and manufacturing, where missing a small defect can lead to significant problems later on.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.