BP Neural Networks with Improved Activation Function and Its Application in the Micrographs Classification

A new activation function employing four adjustable parameters based on the standard Sigmoidal function is put forward. The activation function can adjust the step, position and mapping scope simultaneously, so it has a stronger nonlinear mapping capabilities. Learning algorithm of BP neural networks is also deduced. The simulation results show that comparing with the traditional standard Sigmoidal function, the improved activation function increase the convergence speed more than 10 times while the convergence error less than 1%. It also can reduce the hidden layers´ nodes effectively. Their learning ability can be improved greatly. The efficiency and advantage of the method is proved by the classification results for the Chinese wines´ micrographs based on the improved and traditional BP ANNs.