A neural network model with sparsity constraint and a complete moment feature set are proposed for improving the performance of infrared dim target detection.The traditional neural network models employ such activation functions as SoftMax or Logistic regression for classification.A simpler sign function is designed as the activation function of the classification layer, and forward regression is used to estimate the parameters.In order to reduce computational complexity and improve detection performance, norm constraints are added to the objective function, which can maintain the consistency and sparsity of the parameters.Experimental results show that the new method outperforms traditional approaches and realizes real-time implementation.