TY  - JOUR
AU  - Ali, Muhammad Hamza
AU  - Ali, Farhan
AU  - Muhammad, Zaid
PY  - 2026
DA  - 2026/09/17
TI  - FocusNet: Feature Oriented Contextual Understanding via Bidirectional Supervision for Surface Defect Detection
JO  - ICCK Transactions on Sensing, Communication, and Control
T2  - ICCK Transactions on Sensing, Communication, and Control
JF  - ICCK Transactions on Sensing, Communication, and Control
VL  - 3
IS  - 3
SP  - 160
EP  - 175
DO  - 10.62762/TSCC.2026.978086
UR  - https://www.icck.org/article/abs/TSCC.2026.978086
KW  - surface defect segmentation
KW  - metallic tile inspection
KW  - hybrid attention mechanism
KW  - bidirectional feature pyramid network
KW  - deep supervision
KW  - industrial quality control
AB  - Surface defect segmentation in metallic materials presents significant challenges due to irregular defect shapes, extreme scale variations ranging from small blowholes to large uneven regions, and low contrast with complex background textures. While Convolutional Neural Networks excel at local feature extraction, their limited receptive fields hinder effective global context modeling. Conversely, Vision Transformers capture long-range dependencies but struggle with fine-grained boundary details critical for accurate defect localization. To address these limitations, we propose FocusNet, a novel architecture integrating multi-scale feature refinement, hybrid attention mechanisms, and progressive decoding with bidirectional supervision for robust metallic surface defect segmentation. FocusNet introduces a Feature Refinement Module combining Residual Feature Enhancement Blocks and Atrous Spatial Pyramid Pooling to capture multi-scale contextual information while preserving fine-grained details. A Hybrid Attention Mechanism synergistically fuses channel, spatial, and coordinate attention to enhance discriminative feature representations across both global and local scales. The Bidirectional Feature Pyramid Network enables efficient weighted multi-scale feature fusion, facilitating robust detection across diverse defect characteristics. Our progressive decoder employs deep supervision at multiple hierarchical levels, enforcing intermediate representations to learn defect patterns and improving boundary localization accuracy. Extensive experiments on the MT-Defect benchmark demonstrate state-of-the-art performance with 81.6% mIoU, surpassing LGGFormer (80.0%) and SegFormer (66.1%). Notably, FocusNet achieves particularly strong gains on challenging defect categories, including Crack (70.7% vs. 64.9%) and Break (74.3% vs. 72.4%) compared with the best-performing prior method LGGFormer, demonstrating robust segmentation across defects with diverse scales, orientations, and boundary complexities.
SN  - 3068-9287
PB  - Institute of Central Computation and Knowledge
LA  - English
ER  - 

@article{Ali2026FocusNet,
  author = {Muhammad Hamza Ali and Farhan Ali and Zaid Muhammad},
  title = {FocusNet: Feature Oriented Contextual Understanding via Bidirectional Supervision for Surface Defect Detection},
  journal = {ICCK Transactions on Sensing, Communication, and Control},
  year = {2026},
  volume = {3},
  number = {3},
  pages = {160-175},
  doi = {10.62762/TSCC.2026.978086},
  url = {https://www.icck.org/article/abs/TSCC.2026.978086},
  abstract = {Surface defect segmentation in metallic materials presents significant challenges due to irregular defect shapes, extreme scale variations ranging from small blowholes to large uneven regions, and low contrast with complex background textures. While Convolutional Neural Networks excel at local feature extraction, their limited receptive fields hinder effective global context modeling. Conversely, Vision Transformers capture long-range dependencies but struggle with fine-grained boundary details critical for accurate defect localization. To address these limitations, we propose FocusNet, a novel architecture integrating multi-scale feature refinement, hybrid attention mechanisms, and progressive decoding with bidirectional supervision for robust metallic surface defect segmentation. FocusNet introduces a Feature Refinement Module combining Residual Feature Enhancement Blocks and Atrous Spatial Pyramid Pooling to capture multi-scale contextual information while preserving fine-grained details. A Hybrid Attention Mechanism synergistically fuses channel, spatial, and coordinate attention to enhance discriminative feature representations across both global and local scales. The Bidirectional Feature Pyramid Network enables efficient weighted multi-scale feature fusion, facilitating robust detection across diverse defect characteristics. Our progressive decoder employs deep supervision at multiple hierarchical levels, enforcing intermediate representations to learn defect patterns and improving boundary localization accuracy. Extensive experiments on the MT-Defect benchmark demonstrate state-of-the-art performance with 81.6\% mIoU, surpassing LGGFormer (80.0\%) and SegFormer (66.1\%). Notably, FocusNet achieves particularly strong gains on challenging defect categories, including Crack (70.7\% vs. 64.9\%) and Break (74.3\% vs. 72.4\%) compared with the best-performing prior method LGGFormer, demonstrating robust segmentation across defects with diverse scales, orientations, and boundary complexities.},
  keywords = {surface defect segmentation, metallic tile inspection, hybrid attention mechanism, bidirectional feature pyramid network, deep supervision, industrial quality control},
  issn = {3068-9287},
  publisher = {Institute of Central Computation and Knowledge}
}