There are some usual methods in vehicle license plate location, such as segmentation in grey-level image, color image edge extraction and neural networks filters etc. All these methods are proved not quite satisfactory in various conditions, or are influenced by some factors. In this paper, we present to classify colors of pixels by using improved neural networks, which include 27 nodes of input layer, 30 nodes of hidden layer and 6 nodes of output layer. Several candidate plate regions are extracted from the results of classification. Then a criterion including the features of areas, the ratios of width to height and vertical projection histogram is proposed to decide a real license plate region. Experimental results show that this method has a high locating rate, and adapts to various conditions.