The existing scene text detection algorithms will include redundant background area when dealing with arbitrary shape text regions. In order to automatically learn to focus on arbitrary shape text regions, a text detection algorithm is proposed under attention supervision strategy. Firstly, deep residual network is used as the skeleton network to extract feature maps containing multi-scale information, and then the fused feature maps are converted to generate attention masks through the attention mask generation module, and through the background suppression module, the next level feature map is generated by attention mask. Finally, the segmentation masks are generated through a series of convolution operations, and the final text detection results are obtained after post-processing optimization. Experimental results show that the proposed algorithm has superior multi-index comprehensive performance on the ICDAR2015 dataset, and the <italic>F</italic>-measure is 2.1% ahead of the comparison algorithm.