DFM-DETR: frequency-suppressed Mamba detector for high-reflectivity ceramic plates defects

Surface defect detection is critical in quality control for ceramic plates production. However, the multi-scale characteristics of defects, coupled with highly reflective surfaces and diverse stylistic designs, pose significant challenges to precise localisation and classification. To address these issues, this study proposes a novel object detection paradigm (DFM-DETR). First, a Frequency Window Specular Mitigation Attention (FWSMA) is developed to suppress occlusion effects caused by strong specular reflections or shadows, thereby enhancing detection robustness under complex lighting conditions. In addition, we propose the Dynamic Large-kernel Decomposed Mixer Block (DLDMB), which enables efficient feature extraction by dynamically adapting multi-scale dilated convolution weights and integrating edge-aware enhancement with feature calibration. Finally, a CSP Scan Mamba Module (CSMM) is proposed to introduce long-range dependency modelling capability, aiming to enhance feature representation efficiency and global perception performance in real-time object detection models. Experimental results confirm our model’s robustness in handling specular reflections, small defects on curved components, and complex surface geometries, establishing it as a reliable solution for industrial inspection.