Bridging the Species Gap: VLM-Enhanced Asymmetric LoRA Diffusion for Industrial Defect Detection
Year: 2026
Type: article
Abstract: Deep learning-based wood defect detection systems suffer significant performance degradation when switching between species due to domain shift. Unsupervised domain adaptation (UDA) can mitigate this issue, but its effectiveness is limited when only scarce unlabeled target domain samples are available. In this article, we propose vision language model (VLM)-enhanced asymmetric low-rank adaptation (LoRA) diffusion (VALD), a generative augmentation framework that synthesizes target style defect samples to bridge the species gap. VALD employs a two-phase asymmetric LoRA fine-tuning strategy for stable diffusion. Specifically, the first phase trains a full LoRA with class-specific textual inversion tokens in the source domain; the second phase freezes the downprojection matrix and token embeddings, fine-tuning only the upprojection matrix in the target domain to adapt the visual style while preserving defect semantics.In addition, a large VLM is introduced to extract multidimensional visual background descriptions, which are incorporated into generation prompts to enhance sample diversity. Extensive experiments on 12 cross-species wood defect detection transfer tasks demonstrate that VALD consistently improves multiple UDA methods, achieving up to 90.55% accuracy and outperforming state-of-the-art diffusion-based augmentation approaches.
Institutions: Fuzhou University, University of Hong Kong
Language: English
Subfield: Industrial and Manufacturing Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $2,645
Award: G3254050399
OpenAlex ID: W7207779782