材料科学
电子工程
计算机科学
信号处理
工程类
故障检测与隔离
电气工程
电容
光电子学
噪音(视频)
瞬态分析
探测器
集成电路封装
作者
Yu Zhao,Jing Sun,Guohui Zhang,Fuming Sun,Haojie Li
标识
DOI:10.1109/tim.2026.3693442
摘要
Surface defect detection plays an essential role in modern industrial quality control. Adapting vision foundation models for this task aims to deliver detection capabilities characterized by high accuracy and strong generalization. While models such as the Segment Anything Model (SAM) exhibit great potential in generic segmentation, their effectiveness diminishes in industrial scenarios involving complex backgrounds and low-contrast defects due to insufficient feature perception and imprecise localization. To overcome these challenges, we introduce SDDNet, an efficient framework specifically tailored for industrial defect detection. By incorporating a functionally decoupled dual-adapter strategy, our method adapts the foundation model’s representation capabilities while fine-tuning only 1.8% of the parameters. The core components include a multi-scale perceiver integrating a Shared Transformer Block (STB) and an adaptive decoder dedicated to precise boundary recovery. This approach achieves competitive performance across multiple benchmarks with minimal computational overhead. Experimental results demonstrate that SDDNet compares favorably with the state-of-the-art method SAM2-UNet on the SD-SALIENCY-900 and NEU-RSDDS-AUG datasets. Code and results will be available at https://github.com/wellMachine/SDDNet/.
科研通智能强力驱动
Strongly Powered by AbleSci AI