Detection of Dense, Overlapping, Geometric Objects

Abstract

Using a unique data collection, we are able to study the detection of dense geometric objects in image data where object density, clarity, and size vary. The data is a large set of black and white images of scatterplots, taken from journals reporting thermophysical property data of metal systems, whose plot points are represented primarily by circles, triangles, and squares. We built a highly accurate single class U-Net convolutional neural network model to identify 97 % of image objects in a defined set of test images, locating the centers of the objects to within a few pixels of the correct locations. We found an optimal way in which to mark our training data masks to achieve this level of accuracy. The optimal markings for object classification, however, required more information in the masks to identify particular types of geometries. We show a range of different patterns used to mark the training data masks, and how they help or hurt our dual goals of location and classification. Altering the annotations in the segmentation masks can increase both the accuracy of object classification and localization on the plots, more than other factors such as adding loss terms to the network calculations. However, localization of the plot points and classification of the geometric objects require different optimal training data.