Industrial Wood Defect Detection Using Hybrid Transformer-Mamba-Based Image Super Resolution and Deep Learning

Lai, Zhao Yang (2026) Industrial Wood Defect Detection Using Hybrid Transformer-Mamba-Based Image Super Resolution and Deep Learning. Final Year Project (Bachelor), Tunku Abdul Rahman University of Management and Technology.

Text LAI ZHAO YANG_Full Text.pdf Restricted to Registered users only Download (13MB)

Abstract

Industrial wood inspection faces a practical challenge where deploying high-resolution cameras across all inspection points is cost-prohibitive, leading manufacturers to use lower-cost low-resolution cameras that reduce the visibility of subtle surface defects such as cracks, dents, and gesso lines. Existing CNN and GAN-based super-resolution models improve image clarity but suffer from oversmoothing and artifacts, while transformer-based models offer better reconstruction quality at high computational cost. This study proposes a hybrid Transformer-Mamba super-resolution model derived from the HiT-SRF architecture as a preprocessing step for industrial wood defect detection. Mamba-2 state space blocks replace selected windowed self-attention modules in the deeper stages of the architecture, reducing computational cost while maintaining competitive reconstruction quality. A three-stage ablation study determines the optimal placement pattern, resulting in the final D3 model which achieves a 25.7% runtime reduction over the HiT-SRF baseline while surpassing it in PSNR under identical training conditions. After fine-tuning on the wood defect dataset, D3 achieves the highest PSNR of 47.55 dB among all compared models. A cross-domain detection experiment using YOLOv12m confirms that SR preprocessed images achieve mAP50 of 0.661, nearly identical to the high-resolution reference at 0.662, while direct low-resolution detection drops to 0.445. This validates the proposed framework as a practical solution for bridging the domain gap between low-cost deployment cameras and high-resolution trained detectors without requiring detector retraining. Moreover, the research corresponds to multiple United Nations Sustainable Development Goals (SDGs), especially SDG 9 (Industry, Innovation and Infrastructure), SDG 12 (Responsible Consumption and Production), SDG 15 (Life on Land), and SDG 8 (Decent Work and Economic Growth), as it advances industrial modernization, waste reduction and sustainable forestry