<p>Recently, reconstruction-based defect detection methods have achieved significant performance improvements, but their effectiveness is highly dependent on the quality of the reconstructed images. Many approaches struggle to simultaneously preserve both macroscopic structure and fine-grained details, which can lead to structural errors, or blurred/distorted details even when the overall structure is correct. In this paper, we propose a feature-reconstruction-based defect detection method that employs a Latent Diffusion Model (LDM) as a robust reconstruction baseline. It also incorporates a dual-branch feature-guided network that flexibly integrates attention-layer features through cross-layer feature fusion and cascaded shallow representations, embedding full-frequency image information into the reverse diffusion process to produce high-fidelity reconstructed images. Anomaly detection maps are then generated based on multi-scale feature differences between the input and reconstructed images. We evaluated our method on the MVTec-AD and Visa datasets, achieving image-level AUROC/F1 of 97.0%/96.3% and 87.4%/85.2%, pixel-level AUROC/F1 of 96.7%/56.5% and 95.8%/34.1%, and PRO of 89.6% and 75.9%. Tests on self-built soybean and red jujube datasets yielded image-level AUROC/F1 of 100%/99.8% and 93.0%/90.5%, pixel-level AUROC/F1 of 95.5%/41.5% and 92.6%/30.0%. The experimental results indicate that our method achieves excellent detection accuracy and robustness, demonstrating broad applicability and practical engineering value in defect detection tasks.</p>