Word Recognition in Captured Images by CNN Trained with Synthetic Images

Problems like robotic navigation and automatic geocoding of businesses require artificial agents to perform rapid and accurate word recognition in natural images. We set out to develop a deep learning method to recognize words from different languages in captured images, with high accuracy and with small number of captured samples. Our experiments reveal three main findings. First, we feed images of words as inputs to the neural network directly, omitting segmentation and postprocessing step to avoid compound errors. We found this method to work well for our samples. Second, we are able to train machine learning models to recognize words using purely synthetic training samples by applying feature extractions to both training and testing datasets prior to passing them through deep networks. This achievement allows us to train neural network cheaply on synthetic data and transfer knowledge to recognize words in real data. Third, we set up experiments to compare model performances when using Canny edge detection and Chu's 3D thinning algorithm as preprocessing methods. We found that Canny edge detection performs better in most cases.