In scene text detection, the effective fusion of the detail information and the semantic information is crucial for the accurate results. However, the upsampling operation used in existing methods during the fusion process blurs the precise information carried by feature maps at each scale, and the inference speed of convolution is not optimal. To address this problem, this paper proposes the Refinement Correction Implicit Fusion Network (RCIFNet). In the feature fusion process, the Implicit Feature Fusion Function designed by us is used to obtain the neighboring feature vectors and the corresponding coordinates according to the given query coordinates in the multiscale feature, and input them into Lightweight MLP to realize the implicit feature fusion of multiscale features. Experiments are conducted on four benchmark datasets, and the results verify the superiority of the proposed model compared with current techniques.