DEFectDiffuser: A Text-Free Framework for Semantically Consistent PCB Defect Image Generation via Dual Diffusion Alignment

The scarcity of defective samples in industrial inspection significantly limits the ability of deep learning models to generalize and perform robustly, especially in printed circuit board (PCB) defect detection, where diverse defect morphologies coexist with complex background textures. Ensuring spatial consistency between defects and local backgrounds is essential. To address this, we propose DEFectDiffuser, a text-free conditional generation framework for generating semantically consistent PCB defect images. Our approach integrates a low-resolution defect diffusion model, trained on limited defect samples, with a high-resolution texture reconstruction model that leverages abundant defect-free samples to generate specified defects in designated areas while aligning them with background textures. We optimize local semantics using the DDIM inverse diffusion principle to guide defect generation, incorporating semantic constraints via a local defect diffusion model. Additionally, a dynamically optimized weight mask enables progressive feature fusion between defective regions and backgrounds, enhancing realism and quality. Experimental results show that models trained on our generated samples significantly outperform baselines in defect classification and detection tasks. Compared to existing diffusion-based inpainting methods, DEFectDiffuser achieves superior visual fidelity and semantic consistency, offering a novel and effective solution for improving PCB defect detection.