Screenshots of bank fault often exit in natural scenes. If the text can be accurately detected in the screenshots, it will be able to improve the accuracy of text recognition and improve the case base search and active operation and maintenance capabilities. In order to improve the efficiency of the text detection of natural scenes, an algorithm based on deep learning in natural scene is proposed. Firstly, candidate letters are extracted from the maximum stable extreme region, and candidate texts are generated by single-link hierarchical clustering, then the algorithm makes median filter for the candidate text. Lastly, non-texts are removed by the deep confidence network DBN. Experimental results show that DBN-based approach can effectively improve the accuracy of the text detection of natural scenes, and has better results than traditional methods.