Scene text detection using adaptive color reduction, adjacent character model and hybrid verification strategy

Text detection is a primary task for text recognition and understanding, which can be used in many image analysis techniques. In this paper, we propose an effective scene text detection method including three major steps: connected components (CCs) extraction, character-linking and text/non-text classification. First, for CCs extraction, we design an adaptive color reduction scheme by analyzing image color histogram, which reasonably selects color centers and generates unfixed number of color layers for images in different color complexities. Then, for character-linking, an adjacent character model is built by training an extreme learning machine (ELM), instead of setting various thresholds in previous approaches. Finally, a hybrid text verification strategy is adopted, combining convolutional neural network with ELM for text/non-text classification and performing better than just using one of them. Experimental results on some publicly available datasets illustrate the effectiveness of our method and comparative results with some state-of-the-art algorithms demonstrate our competitiveness.