Spectral-spatial co-learning with adaptive multi-scale calibration for industrial coating defect detection

Surface defect detection plays a critical role in modern manufacturing quality control systems. However, existing deep learning methods face significant challenges in industrial coating inspection: texture-type defects manifest as subtle anomalies that are difficult to distinguish from complex reflective backgrounds, while the multi-scale variations of defects require adaptive feature fusion strategies beyond fixed-weight approaches. Moreover, the lack of publicly available coating-specific datasets has hindered the development of tailored detection algorithms for coating processes. To address these limitations, we propose FreqDETR, a novel framework that systematically integrates frequency-domain enhancement mechanisms with adaptive multi-scale calibration, and contribute a specialized benchmark for coating defect detection. Specifically, this work makes three main contributions: (1) We design the Spectral-Spatial Feature Enhancement Stage (SSFES) that embeds three synergistic frequency-domain modules into the backbone network, capturing textural anomalies and geometric discontinuities through spectral analysis. (2) We construct ICD-4, the first publicly available industrial coating defect dataset, containing 3,099 annotated instances across four representative defect types collected from real-world coating production lines, filling a critical gap in coating quality inspection research. (3) We propose the Adaptive Multi-scale Feature Calibration (AMFC) module that dynamically adjusts the contribution of each scale through cross-scale attention interaction and adaptive weight prediction. Extensive experiments demonstrate state-of-the-art performance: FreqDETR achieves 94.1% AP₅₀ on ICD-4, 77.3% AP₅₀ on NEU-DET, 97.9% AP₅₀ on PCB, and 63.1% AP₅₀ on CSDD, outperforming existing methods while maintaining real-time inference capability. This work demonstrates that frequency-spatial co-learning provides more discriminative feature representations for industrial defect detection.