In　the　past　few　years,　text　detection　in　natural　scenes　has　attracted　increasing　attention　due　to　many　real-world　applications.　Most　existing　methods　only　detect　horizontal　or　nearly　horizontal　texts　and　have　complicated　processes.　When　using　the　neural　network　to　detect　text　in　the　image,　some　ambiguity　and　small　words　are　easy　to　be　ignored　because　of　many　pooling　operations.　Therefore,　this　paper　proposes　an　end-to-end　trainable　neural　network　for　detecting　multi-oriented　text　lines　or　words　in　natural　scene　images.　The　network　fuses　multi-level　features　and　is　guided　by　deep　supervision　during　training.　In　this　way,　richer　hierarchical　representations　can　be　learned　automatically.　The　network　makes　two　kinds　of　predictions:　text/no　text　classification　and　location　regression,　thus　we　can　directly　locate　multi-oriented　words　or　text　lines　without　other　unnecessary　intermediate　steps.　Experimental　results　on　the　ICDAR　2015　datasets　and　MSRA-TD500　datasets　have　proven　that　the　proposed　method　outperforms　the　state-of-the-art　methods　by　a　noticeable　margin　on　F-score.