A novel channel pruning method for deep neural network compression