DCRS: Attention-Guided Dual-Channel Reasoning Segmentation for Industrial Anomaly Detection

IF 17.4 1区 计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE
Ruiyang Jin, Jingjie Zeng, Huayang Li, Zewen Bai, Hongfei Lin, Liang Yang
{"title":"DCRS: Attention-Guided Dual-Channel Reasoning Segmentation for Industrial Anomaly Detection","authors":"Ruiyang Jin, Jingjie Zeng, Huayang Li, Zewen Bai, Hongfei Lin, Liang Yang","doi":"10.1016/j.inffus.2026.104927","DOIUrl":null,"url":null,"abstract":"Industrial Anomaly Detection (IAD) serves as a core component in intelligent manufacturing, requiring accurate pixel-level localization of anomalies. Recently, Multimodal Large Language Models (MLLMs) have introduced semantic interaction capabilities into this field, enabling inspection systems to process natural language instructions. However, most vision-language integration solutions adopt a decoupled “perception–reasoning” paradigm, resulting in error propagation across stages, computational redundancy, and loss of spatial details required for boundary-aware localization. To address this, we propose DCRS: an attention-guided MLLM-based end-to-end reasoning segmentation framework. At its core, DCRS uses the generated <ce:monospace>&lt;SEG&gt;</ce:monospace> position as a shared cross-layer readout anchor and establishes a segmentation-intent-aligned explicit–implicit dual-channel interface. The explicit channel reads the final-layer <ce:monospace>&lt;SEG&gt;</ce:monospace> hidden state as a semantic target specification and maps it to SAM’s sparse prompt space, whereas the implicit channel mines transition-zone <ce:monospace>&lt;SEG&gt;</ce:monospace>-to-image attention and reshapes it into a continuous dense prompt. By coordinating these heterogeneous readouts before mask decoding, DCRS provides fine-grained spatial guidance without introducing a separate anomaly-map generation stage. Experiments on MVTec-AD and VisA demonstrate that DCRS achieves strong anomaly detection and pixel-level localization performance, together with effective synthetic-to-real transfer, consistent cross-dataset generalization, and stable behavior across backbone scales. Evaluation on MMAD further shows that DCRS retains effective industrial diagnostic understanding. Meanwhile, under the same inference setup, DCRS achieves a substantial observed end-to-end speedup over AnomalyGPT, demonstrating the favorable system-level inference efficiency of its unified reasoning-and-segmentation pipeline.","PeriodicalId":50367,"journal":{"name":"Information Fusion","volume":"13 1","pages":""},"PeriodicalIF":17.4000,"publicationDate":"2026-11-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Information Fusion","FirstCategoryId":"94","ListUrlMain":"https://doi.org/10.1016/j.inffus.2026.104927","RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE","Score":null,"Total":0}
引用次数: 0

Abstract

Industrial Anomaly Detection (IAD) serves as a core component in intelligent manufacturing, requiring accurate pixel-level localization of anomalies. Recently, Multimodal Large Language Models (MLLMs) have introduced semantic interaction capabilities into this field, enabling inspection systems to process natural language instructions. However, most vision-language integration solutions adopt a decoupled “perception–reasoning” paradigm, resulting in error propagation across stages, computational redundancy, and loss of spatial details required for boundary-aware localization. To address this, we propose DCRS: an attention-guided MLLM-based end-to-end reasoning segmentation framework. At its core, DCRS uses the generated <SEG> position as a shared cross-layer readout anchor and establishes a segmentation-intent-aligned explicit–implicit dual-channel interface. The explicit channel reads the final-layer <SEG> hidden state as a semantic target specification and maps it to SAM’s sparse prompt space, whereas the implicit channel mines transition-zone <SEG>-to-image attention and reshapes it into a continuous dense prompt. By coordinating these heterogeneous readouts before mask decoding, DCRS provides fine-grained spatial guidance without introducing a separate anomaly-map generation stage. Experiments on MVTec-AD and VisA demonstrate that DCRS achieves strong anomaly detection and pixel-level localization performance, together with effective synthetic-to-real transfer, consistent cross-dataset generalization, and stable behavior across backbone scales. Evaluation on MMAD further shows that DCRS retains effective industrial diagnostic understanding. Meanwhile, under the same inference setup, DCRS achieves a substantial observed end-to-end speedup over AnomalyGPT, demonstrating the favorable system-level inference efficiency of its unified reasoning-and-segmentation pipeline.
DCRS:用于工业异常检测的注意力引导双通道推理分割
工业异常检测(Industrial Anomaly Detection, IAD)是智能制造的核心组成部分,需要对异常进行精确的像素级定位。最近,多模态大型语言模型(Multimodal Large Language Models, mllm)将语义交互功能引入该领域,使检测系统能够处理自然语言指令。然而,大多数视觉语言集成解决方案采用解耦的“感知-推理”范式,导致错误跨阶段传播、计算冗余和丢失边界感知定位所需的空间细节。为了解决这个问题,我们提出了DCRS:一个基于注意力引导的基于mllm的端到端推理分割框架。DCRS的核心是使用生成的SEG>;位置作为共享的跨层读出锚点,并建立分段意图对齐的显式-隐式双通道接口。显式通道将最后一层的隐藏状态作为语义目标规范读取,并将其映射到SAM的稀疏提示空间,而隐式通道挖掘过渡区对图像的注意,并将其重塑为连续的密集提示。通过在掩码解码之前协调这些异构读出,DCRS提供了细粒度的空间引导,而不引入单独的异常图生成阶段。在MVTec-AD和VisA上的实验表明,DCRS具有较强的异常检测和像素级定位性能,并且具有有效的从合成到真实的转换、一致的跨数据集泛化和跨骨干尺度稳定的行为。对MMAD的评价进一步表明,DCRS保留了有效的工业诊断理解。同时,在相同的推理设置下,DCRS比AnomalyGPT实现了可观的端到端加速,证明了其统一的推理和分割管道具有良好的系统级推理效率。
本文章由计算机程序翻译,如有差异,请以英文原文为准。
求助全文
约1分钟内获得全文 求助全文
来源期刊
Information Fusion
Information Fusion 工程技术-计算机:理论方法
CiteScore
33.20
自引率
4.30%
发文量
161
审稿时长
7.9 months
期刊介绍: Information Fusion serves as a central platform for showcasing advancements in multi-sensor, multi-source, multi-process information fusion, fostering collaboration among diverse disciplines driving its progress. It is the leading outlet for sharing research and development in this field, focusing on architectures, algorithms, and applications. Papers dealing with fundamental theoretical analyses as well as those demonstrating their application to real-world problems will be welcome.
×
引用
GB/T 7714-2015
复制
MLA
复制
APA
复制
导出至
BibTeX EndNote RefMan NoteFirst NoteExpress
×
提示
您的信息不完整,为了账户安全,请先补充。
现在去补充
×
提示
您因"违规操作"
具体请查看互助需知
我知道了
×
提示
确定
请完成安全验证×
copy
已复制链接
快去分享给好友吧!
我知道了
右上角分享
点击右上角分享
0
联系我们:info@booksci.cn Book学术提供免费学术资源搜索服务,方便国内外学者检索中英文文献。致力于提供最便捷和优质的服务体验。 Copyright © 2023 布克学术 All rights reserved.
京ICP备2023020795号-1
ghs 京公网安备 11010802042870号
Book学术文献互助
Book学术文献互助群
群 号:604180095
Book学术官方微信
小红书