To improve efficiency of on-line automatic inspecting auto rack girders, we recognize kinds of them in advance. This paper puts forward a vision recognition method based on wavelet transform theory and ART1. Firstly, for the real-time gathered auto rack girders images, we must carry on pretreatment of images about de-noising and enhancing, edge which is enclosed by enhanced images is partitioned to 16 regions, according to information which is position and size of holes in 16 sub-regions, gain two binary character templates which shows numbers, types and distribution of holes; Secondly, study abstract binary code with ART1, and scan up, down weights and corresponding serial number of neuron in database which consists of data of hundreds of auto rack girders. If serial numbers of output neuron are the same which is in two character templates, then this type auto rack girder is inspected. Before on-line inspecting, we gain character templates of different auto rack girders, study them with ART1. Gained the above information are saved in database, they are used as a criterion of recognition. Experiments indicate on-line maximal recognition rate meets demands of production, based on this method, and possessed advantage of more rapid and more precise recognition etc.