JUDO: Pioneering Domain-Oriented AI for Industrial Anomaly Detection

The Critical Need for Smarter Industrial Anomaly Detection

      In the dynamic world of manufacturing and industrial operations, the ability to rapidly and accurately detect anomalies is paramount. From identifying microscopic defects on circuit boards to monitoring critical infrastructure for signs of wear, early detection prevents costly failures, ensures product quality, and enhances safety. Recent advancements in Artificial Intelligence, particularly Large Multimodal Models (LMMs), have shown immense promise in visual anomaly detection by processing diverse human instructions and understanding images with unprecedented depth. However, these powerful general-purpose AIs often face a critical challenge: a lack of specific domain knowledge when confronted with the intricate, specialized environments of industrial settings.

      While LMMs can generalize reasonably well, the true complexity of industrial scenarios demands an AI that not only sees but truly understands the nuances of what constitutes a "normal" versus "defective" component within a highly specific context. This gap often leads to less accurate responses, highlighting the need for solutions that can internalize the unique characteristics of industrial data.

Bridging the Knowledge Gap: Why Generic AI Falls Short in Manufacturing

      Traditional visual anomaly detection methods in industrial settings, often unsupervised, excel at identifying deviations from normal patterns. These methods, including reconstruction-based approaches (like autoencoders) and embedding-based techniques, can pinpoint abnormalities and their locations. However, they typically fall short in providing deeper defect analysis or detailed explanations for their predictions. The advent of LMMs has opened the door for comprehensive defect analysis, moving beyond mere detection to reasoning about the nature and implications of an anomaly.

      Despite these advancements, many current LMMs are primarily optimized for general question-answering tasks. Their inherent knowledge, while broad, is insufficient for the highly specialized visual and textual characteristics of industrial defects. For instance, a subtle discoloration on a specific machine part might be a critical anomaly, but a general LMM may not recognize its significance without contextual understanding. Researchers have explored providing external knowledge or normal samples as context during inference, but this approach has limitations; if the LMM lacks sufficient internalized domain knowledge, it can become overly reliant on external cues, leading to plausible but ultimately inaccurate responses. As highlighted by Kang et al. (2026), this dependency hinders reliable and accurate domain-oriented reasoning.

JUDO's Multi-Stage Approach to Industrial Intelligence

      To address this critical challenge, a novel framework named JUDO (Juxtaposed Domain-Oriented Multimodal Reasoner) has been developed. JUDO is designed to systematically internalize domain knowledge for industrial anomaly detection, moving beyond simple post-training to deep domain alignment across both visual understanding and textual reasoning.

      The JUDO framework operates in three distinct, yet integrated, stages:

• Stage 1: Visual Reasoning through Juxtaposed Segmentation Learning. JUDO initiates its learning by establishing robust visual reasoning capabilities. Instead of treating normal samples as mere optional context, it incorporates them into a core reasoning context during training. This involves a technique called "juxtaposed segmentation learning," where query images (potentially containing defects) are compared side-by-side with known normal images. By performing this visual diffing, the model learns to precisely segment, or delineate, the defect regions. This fine-grained comparative inspection allows the AI to develop a foundational understanding of what deviations look like relative to a perfect state. Solutions like ARSA's AI Video Analytics leverage similar computer vision principles to identify deviations in real-time.
• Stage 2: Enhancing Textual Reasoning with Domain Knowledge. In this stage, JUDO elevates its textual understanding by injecting crucial domain-specific knowledge directly into the model's parameters. This is achieved through "supervised fine-tuning (SFT)," a targeted training process where the AI is exposed to vast amounts of labeled industrial data. This data includes definitions of defects, their potential causes, and consequences, all expressed in textual form. By integrating this information into its core architecture, JUDO builds a foundational internal knowledge base for industrial anomaly reasoning, leading to more reliable and contextually aware explanations. This contrasts with methods that only supply knowledge externally via prompts during inference, ensuring the AI genuinely comprehends the domain.
• Stage 3: Unifying Visual and Semantic Understanding with Reinforcement Learning. The final stage seamlessly unifies the visual grounding learned in Stage 1 with the internalized domain semantics from Stage 2. This is accomplished using "reinforcement learning (GRPO)," an advanced training paradigm where the AI learns through iterative trial and error, guided by a system of "rewards." JUDO employs tailored reward mechanisms designed for domain reasoning accuracy, precise anomaly segmentation, overall choice accuracy, and structural alignment. These rewards incentivize the model to produce accurate, domain-aware anomaly understanding, ensuring its outputs are not only correct but also align with the complex requirements of industrial applications. This integrated learning approach enables the model to connect what it sees with what it knows, leading to powerful and reliable insights.

Practical Impact: Enhanced Reliability, Explainability, and Performance

      The effectiveness of the JUDO approach has been rigorously demonstrated through extensive experiments on the MMAD industrial anomaly detection benchmark. JUDO achieved superior performance compared to leading models like Qwen2.5-VL-7B and GPT-4o. These results underscore the significant importance of enhancing domain knowledge and contextual understanding for effective anomaly reasoning in industrial settings.

      For enterprises, this translates into tangible benefits:

• Increased Accuracy and Reliability: By internalizing domain knowledge and performing visual comparative inspections, JUDO significantly reduces the likelihood of false positives and negatives, leading to more trustworthy detection systems.

Enhanced Explainability: Unlike black-box AI models, JUDO's grounding predictions in localized defect regions through anomaly segmentation and its alignment with domain-specific knowledge offer clear, interpretable insights. This means operators can understand why* a particular area is flagged as anomalous, fostering trust and enabling faster, more informed decision-making.

• Cost Reduction and Efficiency: Automated, accurate anomaly detection minimizes manual inspection time, reduces scrap rates, and prevents downstream failures, all contributing to substantial cost savings.
• Improved Compliance and Safety: In industries where safety and regulatory compliance are critical, precise monitoring of defects (e.g., PPE compliance, restricted area intrusions, product quality) is vital. JUDO's domain-aligned reasoning supports adherence to stringent standards.
• Scalable Deployment: Frameworks like JUDO, emphasizing efficient knowledge integration and robust performance, pave the way for scalable AI solutions that can be deployed across various industrial environments, from individual production lines to large-scale infrastructure monitoring. Such capabilities can be deployed using edge solutions such as ARSA AI Box Series for localized processing and rapid rollout.

ARSA Technology's Role in Real-World AI Deployment

      The principles embodied by JUDO – the deep integration of domain-specific knowledge, multimodal reasoning, and a focus on practical, explainable AI – are central to ARSA Technology’s mission. As an AI & IoT solutions provider, ARSA Technology has been experienced since 2018 in developing and deploying production-ready systems that address mission-critical challenges across various industries. Our approach involves understanding the unique operational realities of each client and engineering tailored solutions that transform complex data into actionable intelligence.

      Whether it's custom computer vision systems for quality control or advanced analytics platforms for operational optimization, ARSA leverages deep expertise to build AI solutions that deliver measurable impact. We bridge advanced AI research with operational reality, ensuring that technology not only works but also generates significant value under real industrial constraints. Our dedication to precision, scalability, and measurable ROI mirrors the cutting-edge advancements seen in research like JUDO, allowing us to implement robust, domain-oriented AI systems for global enterprises.

Conclusion: The Future of Precision Manufacturing and Operations

      The development of frameworks like JUDO represents a significant leap forward in making AI truly effective for complex industrial applications. By systematically internalizing domain knowledge and context, such models can provide highly accurate, reliable, and explainable anomaly detection capabilities. This evolution is crucial for sectors striving for Industry 4.0 automation, where intelligent systems are not just about detecting problems, but understanding their root causes and implications. The ability of AI to combine visual perception with deep domain understanding promises a future of more efficient, safer, and higher-quality industrial operations worldwide.

      To explore how advanced AI and IoT solutions can transform your industrial challenges into intelligent, profitable outcomes, we invite you to contact ARSA for a free consultation.

      Source: Kang, H., Lee, W., Kim, J., & Park, H. (2026). JUDO: A Juxtaposed Domain-Oriented Multimodal Reasoner for Industrial Anomaly QA. Published as a conference paper at ICLR 2026. Available at https://arxiv.org/abs/2605.20284.