Swin Transformer with Feature Pyramid Networks for Scene Text Detection of the Secondary Circuit Cabinet Wiring

The scene text of the secondary circuit cabinet wiring site in the substation includes various bending, occlusion, lighting, and it is difficult to achieve satisfactory detection results using previous approaches. To solve this problem, we propose a method called Swin-FPN, which integrate the advantages of Swin Transformer into the feature pyramid networks (FPN) to effectively improve the performance of FPN. Specifically, we firstly extract the global self-attention contexts of each level of the FPN using Swin Transformer. Then, all levels of the FPN are concatenated by upsampling to produce the feature fusion and project module. Finally, this module is used to predict the score map and keypoints map of the text regions, and thus obtain the final detection result. To evaluate the performance of our method, we construct a scene text dataset with various shapes and occlusion from the substation secondary circuit cabinet wiring site (SSCCWS). We train our Swin-FPN network on public datasets, and then evaluate the performance on our SSCCWS dataset. Experiments demonstrate that the proposed method can achieve better detection performance for SSCCWS scene text compared with state-of-the-art approaches. Thus, our proposed method lays a good foundation for the intelligent inspection of substations.