A high-accuracy segmentation and size measurement method for green pellet images using a deep cascaded network

Green pellets are important material for steel industry. Particle size distribution (PSD) is an important quality parameter of green pellets and is usually measured by a machine vision system installed in the production line. Therefore, segmentation of green pellet images plays a key role in such measurements. Many segmentation algorithms including traditional and deep learning ones are limited in processing images with overlapping pellets and uneven illumination, which greatly affects PSD measuring accuracy. In the presented work, we propose a high-accuracy segmentation method for green pellet images using a deep cascaded network. The cascaded architecture of the network can provide deep and abundant features to the decision layer to improve segmentation accuracy; based on the probability map generated by the network, the shape of each pellet (thus pellet size) can be obtained using contour detection and ellipse fitting. The proposed method was tested by 340 pellet images captured from a steel company, and compared with five current deep networks. Results show that the proposed cascaded network achieves higher segmentation accuracy than other compared networks in terms of f1-score, AUC measures, DSC index (about 0.98, 0.97 and 0.94 in average), and the obtained PSD curve agrees better with ground truth.