A dual-branch cross-attention network for visual inspection of carbon-fiber composite preforms under complex weave textures

Accurate measurement of warp and weft yarn density is essential for evaluating the structural quality of carbon-fiber composite preforms used in aerospace and other high-performance applications. However, automated visual inspection of such materials remains difficult because carbon-fiber images typically exhibit low contrast, local specular reflection, filament hairiness, and weak yarn-boundary continuity. In addition, large-scale annotated datasets are often difficult to obtain in practical industrial inspection, which makes density estimation under limited-data conditions particularly challenging. To address these problems, this study proposes a dual-branch cross-attention framework for automated fabric density inspection of carbon-fiber preforms. The method combines a Transformer-based branch for capturing global weave regularity with a convolutional neural network (CNN)–fast Fourier transform (FFT) branch for strengthening local yarn texture representation, while a cross-attention mechanism enables interaction between global structural information and local feature responses. In addition, a preprocessing procedure integrating nonlocal means denoising, contrast-limited adaptive histogram equalization (CLAHE), Hessian-based ridge enhancement, and Radon-based orientation correction was introduced to improve the visibility and continuity of yarn structures under challenging surface conditions. Under specimen-level fivefold cross-validation, the proposed framework achieved the best overall performance among the tested configurations, with a mean absolute error of 0.332 ± 0.028 and a root mean square error of 0.445 ± 0.036. Compared with the single CNN baseline, the proposed method reduced the mean absolute error by 44.39%. Additional small-sample sensitivity analysis further showed that the proposed framework maintained a relative advantage when the amount of available training data were reduced. Small-sample sensitivity analysis and intermediate feature visualization further indicate that the proposed design reduces overfitting risk by combining global structural constraints with local texture enhancement.