Year: 2026
Type: article
Abstract: Introduction Convolutional neural networks (CNNs) and Transformers have significantly advanced defect detection, improving accuracy across diverse industrial scenarios. However, CNNs struggle to model long-range dependencies, while Transformers often require substantial computational resources for dense, pixel-level defect-detection tasks. To address these challenges, we propose the Multi-Information Perception-Enhanced Mamba UNet (MPMNet). Methods MPMNet uses Visual State Space (VSS) blocks as core components in both the encoder and decoder. To further refine feature representations, the Residual Attention (RA) module suppresses redundant or irrelevant information through combined channel and spatial attention mechanisms. In addition, the Split Context Interaction (SCI) module processes low-level features with dilated convolutions to enhance detail extraction, while the Semantic-Aware Interaction (SI) module optimizes location cues in high-level features using attention mechanisms. Finally, cross-level connections during decoding mitigate information dilution. Results Experiments on four public defect datasets demonstrate that the proposed MPMNet achieves competitive results. Discussion These designs collectively enable precise and efficient defect detection by enhancing multi-information perception, refining feature representations, and mitigating information dilution during decoding.
Authors: Dianlong Chen, Tengbin Liu, Zisheng Xu, Chao Tao
Language: English
Subfield: Industrial and Manufacturing Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,353
OpenAlex ID: W7236860049
