The knot on the wood surface is a very important kind of wood defects, and it is the key specification for assessing the appearance grade and the quality of lumber and veneer. To enhance the accuracy and efficiency of knot defects recognition, and improve the automatic level of detecting procedure, the recognition of knot defects by using the statistics features of gray-scale histogram from wood surface image is studied. The classifying ability of seven statistics features is evaluated through using the between-cluster distance, and hence the optimal statistics feature that recognizes the knot defect is determined, such as the smoothness. At the same time, an adaptive clustering method with maximal between-cluster variance is presented to determine the classifying threshold, and then based on that the knot defect is recognized. The online detection experiment shows that the recognition rate of the presented method is up to 99%.