From the viewpoint of real-time application, the MADALINE network for invariant feature extraction is improved. Presented are the algorithm structure for the MADALINE invariant network with local interconnection, which is simplified to usual matrix computation, and the approaches by learning to determine the adaptive MAJ threshold and the fixed MAJ threshold. The improved network with local interconnection is implemented for binary patterns. The comparisons of MADALINE with its improved version are made.