TLDR: AutoNAD is an automated framework for industrial surface defect detection that combines convolutions, transformers, and MLPs to handle diverse defect appearances and similarities. It features a cross weight sharing strategy for efficient training, a searchable multi-level feature aggregation module for multi-scale feature learning, and a latency-aware prior for efficient deployment. Validated on multiple datasets and a real-world platform, AutoNAD achieves high accuracy and efficiency, making it suitable for industrial automation.
Ensuring product quality and manufacturing reliability is a cornerstone of industrial success. A critical aspect of this is industrial surface defect detection (SDD), which involves identifying flaws on product surfaces. However, this task is notoriously challenging due to two main issues: defects of the same type can look very different (intraclass difference), and different types of defects can appear remarkably similar (interclass similarity). Traditionally, engineers have relied on manually designed models, a process that is time-consuming, prone to trial and error, and often falls short in addressing both these complexities effectively.
A new framework called AutoNAD (Automated Neural Architecture Design) has been proposed to tackle these challenges. AutoNAD offers an automated approach to designing neural networks specifically for SDD. Its core innovation lies in its hybrid architecture, which intelligently combines three powerful types of neural network components: convolutions, transformers, and multi-layer perceptrons (MLPs). This blend allows the system to simultaneously capture both the fine, localized details of defects and the broader, long-range contextual information, which is essential for accurate detection.
The framework introduces several key advancements to make this automated design process efficient and effective. One such innovation is the ‘cross weight sharing strategy’. In the world of neural architecture search, many potential network designs (subnets) are evaluated within a larger ‘supernet’. This strategy accelerates the training of this supernet and improves the performance of the individual subnets by allowing different types of operators to share weights efficiently. This means that instead of training each component independently, AutoNAD enables a more unified and faster learning process.
Another significant feature is the ‘searchable multi-level feature aggregation module’ (MFAM). Defects can appear at various scales and have different characteristics. The MFAM is designed to automatically learn how to best combine features extracted at different levels of detail, adapting dynamically to the specific defect types and image scales it encounters. This ensures that the model can effectively process both subtle textures and larger structural flaws.
For industrial deployment, speed is as crucial as accuracy. AutoNAD addresses this with a ‘latency-aware prior’. During the training of the supernet, the system collects real-time performance statistics. This data is then used to guide the search algorithm, favoring architectures that not only perform well but also execute quickly on target hardware. This lightweight approach helps select efficient models without needing complex prediction tools.
The effectiveness of AutoNAD has been rigorously tested on three diverse industrial defect datasets: NEU, MSD, and MT. The results demonstrate that AutoNAD achieves superior detection accuracy and efficiency compared to many existing state-of-the-art methods, including those based purely on CNNs, transformers, or MLPs. Interestingly, even large foundation models, often praised for their general vision capabilities, performed relatively poorly on these industrial datasets, highlighting the need for domain-specific solutions like AutoNAD. This is largely because these general models are not trained on industrial images and lack the specialized knowledge required.
Furthermore, AutoNAD exhibits remarkable search efficiency, requiring substantially less time to find optimal architectures compared to other automated design methods. Its models also show excellent generalization, meaning a model trained on one dataset can perform well when applied to other, unseen industrial datasets, which is vital for practical quality inspection.
To validate its real-world applicability, AutoNAD has been integrated into an automated imaging and detection platform for aero-engine blades. This platform, featuring robotic arms, motion platforms, and high-precision cameras, uses AutoNAD to automatically design and deploy detection networks based on collected data. The system achieved high accuracy and fast inference times on actual production line images, even on edge devices like the Jetson Xavier NX. The architecture search is typically a one-time cost, with minimal retraining needed for minor data shifts, making it a highly practical solution for intelligent manufacturing systems.
Also Read:
- Syn-Diag: Smart Fault Diagnosis for Industry with Cloud-Edge AI
- VER: A New Vision Expert Transformer for Adaptive Robot Learning
In conclusion, AutoNAD represents a significant step forward in automated surface defect detection. By intelligently combining different neural network operators, optimizing training efficiency, and incorporating real-world deployment considerations, it offers a robust and adaptable framework for enhancing quality control in industrial settings. For more details, you can refer to the original research paper.
