An incremental approach for the detection of legend text in digital maps

The presented work concerns the automatic detection of legend texts inside maps. After extracting the texts from the images using OCR tools, we use an iterative clustering process on the extracted texts. We consider five main criteria, with different levels of importance: text alignment, distance between text boxes, background color of the text, font color, and font size. For each criterion, we define appropriate similarity measures. We propose a method that combines, incrementally, the partitions obtained by each criterion. The experimental study reveals two important results. First, combining several criteria gives better results than considering a single distance metric (e.g., Euclidean distance) between text boxes. Secondly, the overall effectiveness of the priority relation, which we intuitively defined among the criteria for detecting caption texts, is confirmed.