A frequency-texture-aware vision method for reliable measurement and inspection of concrete surface defects
Abstract
Vision-based concrete inspection estimates defect category, location, spatial extent and confidence from surface images, but repetitive texture, illumination variation, scale changes and feature misalignment undermine measurement reliability. We propose FTC-YOLO, a measurement-error-aware visual-inspection framework implemented with a frequency-texture-aware YOLOv8 detector. Its Frequency-Domain Enhancement Module re-weights spectral features to preserve defect boundaries and suppress low-frequency noise; the Texture-Adaptive Attention Module directs the neck to defect-bearing regions; and the Cross-Scale Alignment Fusion Module aligns deep and shallow features before channel fusion. On a seven-category construction-defect benchmark, FTC-YOLO attains 0.920 ± 0.011 mAP@0.5 and 0.750±0.013 mAP@0.5:0.95 across three seeds, versus 0.860 ± 0.012 and 0.685 ±0.014 for YOLOv8s. On the common correctly matched ground-truth subset, the mean image-domain centre error decreases from 0.01764 (P95 0.02910) to 0.01224 (P95 0.02042). On a 400-image author-collected defect-free hold-out, false positives decrease from 0.10 to 0.07 per image and specificity increases from 0.90 to 0.93. On a disjoint 400-image author-collected positive hold-out with 2,486 annotations, FTC-YOLO achieves 0.905±0.004 mAP@0.5 and 0.737±0.006 mAP@0.5:0.95, versus 0.842 ± 0.004 and 0.663 ±0.006 for YOLOv8s. The model retains at least 0.840 mAP@0.5 under brightness, blur, JPEG and input-size perturbations, and has 13.2M parameters, requires 38.4 GFLOPs for a single forward pass, and achieves an inference speed of 128 FPS at an input resolution of 640 × 640 pixels on a single NVIDIA RTX 5090 GPU (batch size 1, FP16).