Rapid detection of textile fiber components is very important for production process of quality control, trading and market surveillance. The objective of this research was to assess cotton content in cotton/ramie blended fabric quickly by near infrared (NIR) spectrum technology and variable selection methods. Reflectance spectra of samples were acquired by a NIRFlex N-500 Fourier spectroscopy in the range of 4 000～10 000 cm-1, primary election of spectral range and pretreatment analysis were conducted first. Then, three variable selection methods such as UVE (uninformative variables elimination), SPA (successive projections algorithm) and CARS (competitive adaptive reweighted sampling) were used to select sensitive variables. After that, PLS (partial least squares) was used to develop calibration model for cotton content of cotton/ramie blended fabric, and the best calibration model was used to predict cotton content of samples in prediction set. The result indicates that range of 4 052～8 000 cm-1 is optimal spectral range for cotton content modeling. CARS method is an efficient method to improve model performance, the correlation coefficient and root mean square error of CARS-PLS for calibration and prediction sets are 0.903, 0.749 and 8.01%, 12.93%, respectively. So NIR spectra combined with CARS method is feasible for assessing cotton content in cotton/ramie blended fabric, and CARS method can simplify model, improve model performance.