Segmentation-based compound figure detection and separation methods

Figure detection, separation and image classification are common problems occurring in various fields, especially medicine. Since image databases are usually large, manual classification would be a demanding task. In this paper, we proposed a method for automatic compound figure detection and separation, and gave a comparison between other recognition methods, such as convolutional neural networks. The proposed method is based on differentiating objects in the image, and merging object artifacts with the nearest large object. Parameters of size and distance were varied, and different criteria for determining the object boundaries were tested. Using this method, an accuracy of 90.20% was achieved on a test set of 500 images, with an average processing time less than 600ms per image for the given combination of parameters.