<p>We explore the application of transfer learning for handwritten text recognition. Typical transfer learning leverages pretrained convolutional neural network models to extract hierarchical features from images. However, these feature extractors are trained on imageNet, which contains natural photos with completely different characteristics from the target handwritten data. To address this limitation, we present ParallelNet, a novel dual-branch architecture designed to generalize handwritten data. The architecture consists of a generic feature extractor based on transfer learning and a shallow, domain-specific feature extractor fully trained on the target dataset. By incorporating features from the shallow CNN into the fine-tuning process, we extend the traditional transfer learning paradigm to better align the model with the target domain. Our method was evaluated on 5 handwritten digit datasets covering 4 different languages using 4 pretrained CNN models that were further fine-tuned with 3 different classifiers, and the results consistently demonstrated better or equally better accuracies in most cases as compared to standard transfer learning.</p>