FSQR-Former: A Frequency–Spatial Query Refinement Transformer with Feature Feedback for Surface Defect Segmentation

Accurate surface defect segmentation remains challenging when thin and low-contrast defects are embedded in repetitive textures, where local frequency variations may originate from both defects and background structures. This work presents FSQR-Former, a query-based mask-classification transformer that couples frequency–spatial query refinement with cross-stage feature feedback. Content-Decoupled Directional Wavelet Frequency Attention (CDWFA) applies Haar decomposition to visual features without explicitly added positional encodings and combines directional calibration with frequency gating informed by low-frequency context. Multi-Scale Query-Guided Sparse Attention (MS-QSA) enriches spatial features with neighborhood context and selectively aggregates tokens according to query compatibility. Query-to-Feature Feedback Modulation (Q2FM) converts refined query states into channel and spatial modulation signals for finer-scale features, coupling query refinement with feature recalibration during coarse-to-fine decoding. The three-run mean results of FSQR-Former indicate competitive benchmark performance on ESDIs-SOD and CrackSeg9k, with the best scores across all five evaluation metrics on ZJU-Leaper. Controlled ablations support the contributions of joint frequency–spatial refinement and query-conditioned feature feedback to pixel-wise accuracy and structural agreement. Complexity analysis further indicates a favorable balance between segmentation accuracy, computational cost, and parameter count.