Automatic inspection of woven textile fabric is discussed. A two-stage detection process is adopted, with the second stage involving set of novel contextual decision fusion techniques. Three significant problems are addressed: (1) texture feature extraction: Fourier transform features are found to be well matched to the spatially periodic nature of the woven pattern; (2) detection of localized flaw patterns: since prior probabilities are impossible to estimate, and we cannot hope to enumerate all defect classes, a Neyman-Pearson approach is adopted, i.e., flaw detection is via measured deviation from nominal; and (3) detection of extended flaw patterns: the most common flaws are characterized by linear or other cluster shaped patterns; although these are weakly detectable by local detectors, they may be ignored when local detector sensitivity is set to achieve tolerably low false-alarm rates; a localextended contextual decision fusion technique using morphological filtering enables us to achieve very low composite false-alarm rate. The performance of the system is evaluated on samples of denim fabric containing real defects. The predicted composite false-alarm rate is of the order 1 in 1013, or equivalent to 1 per 100 km of fabric roll. Experimental results demonstrate the compatibility of this favorable false-alarm rate with the reliable detection of flaws, which have been chosen for their subtlety and detection difficulty.