To discriminate the ripeness of cotton quickly and accurately, 15 shape structure features were extracted from cotton images and the execute efficiencies and classification accuracy of their feature subset selection algorithms such as Wrapper-based Exhaustive searching and Wrapper-based stopping(WE-W) and Filter-based Heuristic searching and Wrapper-based stopping(FH-W) were compared by using 10-fold cross-validation. By taking the error rate of a Bayes classifier on validation set (WE-W) and the class-separability measuring value on a training set (FH-W)as assessing functions, the optimal l (l=1,2,3, …, 15) feature subset was searched by using exhaustive (WE-W) and heuristic (FH-W) strategies on the training set, which stops at the minimum error rate of Bayes-classifier on the validation set(WE-W and FH-W). Experimental results show that the average classification rates of WE-W and FH-W algorithms on the prediction set are 85.39% (WE-W) and 85.28% (FH-W) at l=3, respectively. It concludes that the FH-W algorithm can be a reference in practice for its higher execute efficiency and good classification accuracy.