<p>With the rapid advancement of convolutional neural networks (CNNs), CNN-based text image super-resolution methods have achieved remarkable progress. However, to enhance text recognition accuracy, most existing methods emphasize enhancing character readability when improving image clarity, but often overlook restoring fine visual details present in the original high-resolution (HR) images. To overcome these limitations, we propose a Multi-Task Collaborative Network for Text Image Super-Resolution and Text Recognition (SR-TRNet). This framework integrates a super-resolution module with a text recognition module through joint training, enabling end-to-end generation of HR images and simultaneous text recognition. To strengthen the network’s representational capacity, we introduce a Visual Dual Flow Integration (VDFI) module. This architecture exploits the complementary strengths of global statistical features and local fitting ability to obtain multi-granularity visual information with contextual understanding. To further improve overall performance, we design a feature-sharing strategy. Specifically, the super-resolution and text recognition modules share a feature extraction network, thereby reducing parameters. Experimental results demonstrate that our method outperforms most existing approaches in both recognition accuracy and image fidelity, while producing visual effects closer to those of the original HR images.</p>