Further work was done on the cascaded neural network system for muti-target recognition. Three plane models were used as targets to be recognized to imitate the real scene. It is found that the recognizing rate is lowered dramatically mostly due to complex distortions occuring in applications, which are very different with that stimulated by computers. Utilizing images of the models as training samples, the network was reconstructed and retrained. Good results were obtained. And to approach image preprocessing problems, such as location and segmentation, an efficient method of quick target detection and location is proposed based on binary morphological erosion algorithm.