Abstract: A feedforword network of 5-layer is presented, Which accomplishes numeral features extraction and recognition. This network employed weight share and partial reflection to reduce the total weight numbers. In trainning process, an extension BP algorithm with the second derivative of cost function is used. The algorithm has nice convergence properties, which performs four or five times faster than the conventional BP algorithm. In addition, the network weight numbers are greatly reduced by using the seeond derivative information for neural damage or robustness test.