Mesoscopic model has been proven to impose crucial influence on predicting the damage and failure behavior of textile composites. In classical mesoscopic models, the section of yarn is usually simplified as circular, lenticular, ellipsoidal or more sophisticated shapes defined by power ellipse. A numeric tool is proposed in this work to reconstruct the yarn section based upon realistic geometry via X-ray micro-tomography techniques. The sectional shape of yarn is fit as polygon by connecting the selected key boundary points with Graham Scan convex hull algorithm. Improved Lubachevsky-Stillinger algorithm is then employed to drive the motion and growth of the fiber sections, which are considered as circles here, within the yarn boundary until the pre-assigned fiber volume fraction is reached. The proposed method shows excellent performance to model the localized fiber distribution. A representative size of the numeric model is determined following by a strength-geometry sensitivity research on different section models. Further investigation is completed to study the influence of fiber-matrix interface on the damage and failure behavior of the yarn. This approach can be used for arbitrary textile composite where the fiber can be taken as circle at microscale, as long as the real geometry of yarn is provided no matter by micro-tomography or other experimental observation techniques.