In the research of channel decoding combined with deep learning technology, the problem of dimension limitation has al- ways been the focus of researchers. Since the deep neural network is storage intensive, the channel decoder of deep neural network usually needs much more computation consumption and memory than the conventional belief-propagation (BP) decoding. In order to alleviate this problem, an improved neural network decoder for LDPC code is proposed. According to the weight parameter value distribution in the deep neural network channel decoder, weight parameters are selectively added to a new neural network decoder. By limiting the number of training parameters, the scale of the deep neural network channel decoder is reduced. And our algorithm gets a large decoding gain than BP decoding.