In order to resolve the contradiction between computing performance and accuracy of the traditional neural network with continuous weights, and its characteristic tidy memory capacity in embedded systems, a neural network optimization method is proposed. Firstly, we represent the weights of neural network with integers and train the neural network using the Genetic Algorithm. Secondly, the continuous nonlinear-activation function of the neuron is transformed into discrete and linear function using the least-squares arithmetic. Then, the optimized neural network is applied to the image sharpening for verifying its feasibility. Results of experiment show that the new method has a good real time capability and effect in hardware.