Abstract: Curved text detection technology in natural scenes is mostly used in intelligent tourism scenes. Due to the influence of the receptive field size and feature extraction ability of convolutional neural network on the current curved text detection, the network is difficult to identify text and non text areas in natural scene images. A text detection method based on attention mechanism and cavity convolution in natural scene (Resnet Squeeze and Exception Diffusion Jacob Progressive Scale Expansion Network RSDJ-PSE) is proposed. RSDJ-PSE introduces the soft attention mechanism SE block into the backbone network of the detection network, which further enhances the feature extraction capability. Then it introduces the hole convolution into the backbone network, which expands the receptive field of the convolution without increasing the number of parameters. Finally, it uses the Jackard coefficient to replace the Dice coefficient in the post-processing algorithm, which improves the F value of the text detection method. The detection results on directional text dataset ICDAR2015, standard curved text dataset CTW1500 and Total Text dataset show that this method has the best text detection performance compared with eight detection methods.