Model and Data Integrated Transfer Learning for Unstructured Map Text Detection

Abstract

The emergence of the third information wave makes extensive maps available to be generated by volunteered ways, never specially designed and generated by professional institutes alone. These large-scale images-based volunteered maps created by the public provide plentiful geographical information regarding a place while posing a challenge for recognizing the unstructured text in these maps for previous approaches to standard map text detection. Map text or map annotations denote the critical element of map content. To achieve the detection of unstructured map text, this paper proposed an integrated data-based and model-based transfer learning model, which mainly respectively included data augmentation techniques and adaptive fine-tuning, to reinforce the state-of-the-art CNNs by transferring the OCR knowledge for detecting the unstructured text units in volunteered maps. The experiment proved that our proposed framework can effectively reinforce the state-of-the-art CNN in detecting unstructured map text. We hope our research results can contribute to unstructured map text detection and recognition.