AE-BiseNet: Anomaly-Enhanced Bilateral Segmentation Network for Industrial Defect Detection

TY - GEN

T1 - AE-BiseNet

T2 - 9th International Workshop on Advanced Computational Intelligence and Intelligent Informatics, IWACIII 2025

AU - Wang, Xinyang

AU - Wu, Xinwei

AU - Hou, Yongjie

AU - Ma, Hongbin

AU - Jin, Ying

N1 - Publisher Copyright: © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2026.

PY - 2026

Y1 - 2026

N2 - Industrial defect detection is of great significance in enhancing the level of manufacturing industries and achieving the high-end development of manufacturing. Segmentation is an important technology to defect defects. While complex and variable backgrounds, subtle and small defects, class imbalance and sparseness make it difficult to accurately identify defects. The article proposes Anomaly-enhanced Bilateral Segmentation Network (AE-BiseNet) to improve segmentation performance by feature contrasting spatial path with spatial defect enhancement auxiliary loss and local positive and negative representation contrasting auxiliary loss. The former creates a dual-branch structure to capture rich low-level spatial features and trained with auxiliary loss to distinguish normal area and defects. The latter extract the representative high-level features of normal area and defects and mitigate class imbalance problem. Both the auxiliary loss function use ground truth to help confirm normal features and defect features. The stamped parts data set is built to validate the model. Experiments show the proposed method reach mean IoU of 83.8% and defect IoU of 68.0% on the stamped parts data set and mean IoU of 88.1% and defect IoU of 79.4%.

AB - Industrial defect detection is of great significance in enhancing the level of manufacturing industries and achieving the high-end development of manufacturing. Segmentation is an important technology to defect defects. While complex and variable backgrounds, subtle and small defects, class imbalance and sparseness make it difficult to accurately identify defects. The article proposes Anomaly-enhanced Bilateral Segmentation Network (AE-BiseNet) to improve segmentation performance by feature contrasting spatial path with spatial defect enhancement auxiliary loss and local positive and negative representation contrasting auxiliary loss. The former creates a dual-branch structure to capture rich low-level spatial features and trained with auxiliary loss to distinguish normal area and defects. The latter extract the representative high-level features of normal area and defects and mitigate class imbalance problem. Both the auxiliary loss function use ground truth to help confirm normal features and defect features. The stamped parts data set is built to validate the model. Experiments show the proposed method reach mean IoU of 83.8% and defect IoU of 68.0% on the stamped parts data set and mean IoU of 88.1% and defect IoU of 79.4%.

KW - AE-BiseNet

KW - ESemantic Segmentation

KW - Industrial Defect Detection

UR - https://www.scopus.com/pages/publications/105030936878

U2 - 10.1007/978-981-95-6730-0_38

DO - 10.1007/978-981-95-6730-0_38

M3 - Conference contribution

AN - SCOPUS:105030936878

SN - 9789819567294

T3 - Communications in Computer and Information Science

SP - 563

EP - 578

BT - Advanced Computational Intelligence and Intelligent Informatics - 9th International Workshop, IWACIII 2025, Proceedings

A2 - Ma, Hongbin

A2 - Xin, Bin

A2 - She, Jinhua

A2 - Dai, Yaping

PB - Springer Science and Business Media Deutschland GmbH

Y2 - 31 October 2025 through 4 November 2025

ER -