Feature point detection based on convolutional neural network (CNN) has been studied widely. The effective approaches for improving detection accuracy are building a deeper network or using a multi-network cascade structure. However, some potential capacity of CNN has not been excavated. In this stu