Based on one of the improved probabilistic neural network models-FDO neural network, the paper presents an idea of taking the effect of the & neighbors of every pixel into consideration when designing the convergent domain of the network. The activation function of the network is modified, thus endowing the network with good stability and high running speed. The recognition capabilities of FDO network and its improved version are compared in the simulation experiment. The result proves that the improved network is especially suitable for the recognition of targets with Gaussian noise.