Quality control (QC) is a critical component of industrial processes that ensures product reliability, quality, and safety. Anomaly detection (AD), which refers to the process of identifying outliers or rare/unusual events compared to the majority, is crucial for identifying defects during product inspection and QC. The increasing stringency in industrial regulations and rising demand for various products call for automated, robust, and efficient AD systems that can accurately detect anomalies. However, AD becomes particularly challenging using traditional methods, given the obscure and diverse environments in industrial settings, including low-light conditions. Moreover, AD models that rely on low-light image enhancement may be limited by artifacts and noisy images that do not accurately reflect subtle defects on industrial surfaces. Additionally, deep learning-based AD systems require extensive data processing and computational resources, which limit their widespread practical application.
To overcome this challenge, Dr. Phan Xuan Tan, an Associate Professor at the Innovative Global Program, College of Engineering, Shibaura Institute of Technology, Japan, along with Dr. Dinh-Cuong Hoang and other researchers from FPT University, Vietnam, have designed DarkAD—a novel end-to-end framework that can enhance AD in low-light industrial environments. The researchers have introduced a Dark-Aware Feature Adapter (DAFA) that integrates noise reduction and low-light image processing.
Giving further insight into their work, Dr. Tan explains, "Unlike existing methods that rely on computationally expensive low-light image enhancement, DarkAD introduces DAFA, which enhances feature extraction through Frequency-Based Feature Enhancement (FFE) to suppress noise and Illumination-Aware Feature Enhancement (IFE) to amplify critical features in poorly lit areas. The proposed feature enhancement approach allows for real-time AD, reducing inspection errors and operational costs."
Their work has been published in Volume 25 on March 1, 2025 in the Results in Engineering and was made available online on February 7, 2025.
Conventional methods based on reconstruction and feature embedding use pre-trained model sets to identify deviations, while synthesizing-based models generate anomalies in normal images to expand the data set. However, these approaches are limited by semantic conflicts, large memory storage requirements, and the inability to accurately mimic surface anomalies. A hybrid approach that combines the strengths of different methods can improve the robustness of AD systems. SimpleNet is a hybrid approach that combines feature-embedding and synthesizing-based strategies, allowing abstract and flexible anomaly generation and computationally efficient AD. Nonetheless, low-light detection continues to remain a concern. The researchers sought to adapt the SimpleNet model to improve AD in low-light and noisy conditions.
In the current framework, the FFE module enhances low-frequency structural features while reducing high-frequency noise, thereby enabling robust AD even in low-light conditions. The IFE module estimates illumination across the image and enhances regions that are poorly lit, thus mitigating challenges that result from uneven illumination. Notably, the DarkAD model does not require pre-processing or enhancement of the input image. Further, dynamic adaptation by the model selectively amplifies features from well-lit regions, while preserving crucial features from low-lit regions, thus improving its detection accuracy.
In addition to designing the AD model, the researchers also assembled an anomaly training dataset using images of industrial objects with diverse shapes, sizes, colors, and materials acquired in low-light settings. They carefully selected objects that would represent commonly encountered industrial items, increasing the real-world applicability of the model. Their dataset included defect-free and defective objects that reflect common anomalies, including scratches, dents, discolorations, missing parts, and surface deformations. Finally, they combined the newly acquired data with existing datasets to enhance the robustness and scope of the model across diverse industrial settings.
The DarkAD model designed in this study significantly outperformed the SimpleNet model by accurately detecting subtle anomalies, even in objects with complex textures in poorly illuminated conditions. The model also achieved high detection speed, consistency, and localization accuracy compared to other state-of-the-art models. Overall, the DarkAD framework is a robust, high-performing, adaptive, and industrially scalable AD model that can be applied in diverse real-world industrial settings. Its accuracy in detecting anomalies of varying sizes and shapes across diverse materials and complex lighting conditions makes it a valuable QC tool for automated industrial manufacturing, infrastructure monitoring, and detection of instrument malfunctioning and other industrial hazards.
Highlighting the diverse applications of their model, Dr. Tan says, "DarkAD can be potentially applied in various applications, for example, manufacturing QC for detecting defects in automotive parts like clutches and tires, industrial components including cable glands and insulators, and textiles under poor lighting. It can also enable automated 24/7 monitoring and close visual inspection for detecting subtle anomalies in low-light factories, warehouses, high-risk settings like power grid systems, and complex underwater environments, thus reducing reliance on human inspectors."
