This paper discusses methods for character extraction on the basis of statistical and structural features of gray_level images,and proposes a dynamic local contrast threshold method accommodating to line width.Precise locating of character string is realized by exploiting horizontal projection and character arrangements of binary images in horizontal and vertical directions respectively.Also discussed is the method for segmentation of characters in binary images,which is based on projection taking stroke width and character sizes into account.A new method for character identification is explored,which is based on compound neural networks.A complex neural network consists of two sub_nets,the first sub_net performs self_association of patterns via 2_dimentional local_connected 3_order networks,the second sub_net,linking with a locally connected BP networks,performs classification.The reliability of the network recognition is reinforced by introducing conditions for identification denial.Experiments confirm that the proposed methods possess the advantages of impressive robustness,rapid processing and high accuracy of identification. This paper discusses methods for character extraction on the basis of statistical and structural features of gray_level images, and proposes a dynamic local contrast threshold method accommodating to line width. Precise locating of character string is realized by exploiting horizontal projection and character arrangements of binary images in horizontal and vertical directions respectively. Also discussed is the method for segmentation of characters in binary images, which is based on projection taking stroke width and character sizes into account. A new method for character identification is explored, which is based on compound neural networks. A complex neural network consists of two sub_nets, the first sub_net performs self_association of patterns via 2_dimentional local_connected 3_order networks, the second sub_net, linking with a locally connected BP networks, performs classification. Reliability of the network recognition is reinforced by introducing conditions for identification denial. Experiments confirm that the proposed methods possess the advantages of impressive robustness, rapid processing and high accuracy of identification.