Double supervision for scene text detection and recognition based on BMINet

Abstract

Scene text detection and recognition currently stand as prominent research areas in computer vision, boasting a broad spectrum of potential applications in fields such as intelligent driving and automated production. Existing mainstream methodologies, however, suffer from notable deficiencies including incomplete text region detection, excessive background noise, and a neglect of simultaneous global information and contextual dependencies. In this study, we introduce BMINet, an innovative scene text detection approach based on boundary fitting, paired with a double-supervised scene text recognition method that incorporates text region correction. The BMINet framework is primarily structured around a boundary fitting module and a multi-scale fusion module. The boundary fitting module samples a specific number of control points equidistantly along the predicted boundary and adjusts their positions to better align the detection box with the text shape. The multi-scale fusion module integrates information from multi-scale feature maps to expand the network’s receptive field. The double-supervised scene text recognition method, incorporating text region correction, integrates the image processing modules for rotating rectangle boxes and binary image segmentation. Additionally, it introduces a correction network to refine text region boundaries. This method integrates recognition techniques based on CTC loss and attention mechanisms, emphasizing texture details and contextual dependencies in text images to enhance network performance through dual supervision. Extensive ablation and comparison experiments confirm the efficacy of the two-stage model in achieving robust detection and recognition outcomes, achieving a recognition accuracy of 80.6% on the Total-Text dataset.