To solve the problem that singular-modal image recognition using only the low-level visual features has low accuracy, considering that many images have embedded-in textual information, a collaborative method using the embedded-in text to aid the recognition of images is proposed. The method includes three steps. Firstly, after localization, segmentation, binarization and feature extraction, semantics of text is gotten. Secondly, the collaborative posterior probability is calculated by extracting visual features of images and counting correlation of visual and textual modals. At last, for each class of images, the joint posterior probability is calculated using the previous two items. A new image is recognized to the class with maximal joint posterior probability. Experiments on the self-built data set of sports video frames showed that the proposed method performed better than the singular-modal method on three different visual features and had higher accuracy.