In this paper, we propose a simple and efficient feature extraction method to improve the accuracy and speed of texture classification. Our approach is based on features boundary and can be used in multiple-classifier approaches. We split an image into some non-overlapped partitions and extract features from the sub-images. Considering all partitions, boundary of features makes a criterion for pre-classification in the first stage of proposed serial multiple classifier system (FB-ED). Euclidean distance is used as similarity measures for second classifier. Well-known Haralick's features for evaluation of our approach are used. To generalize our approach we employ images in the Brodatz and VisTex data sets. Experimental results have depicted FB pre-classifier makes classification more accurate and significantly faster than single stage classification on considered datasets.