Amodal segmentation of cane sugar crystal via deep neural networks

Abstract

The growth of cane sugar crystals can provide fundamental information for the setting of production process parameters. However, the overlap phenomenon of sugar particles has adverse effects on identifying the shape of sugarcane crystals in images, and it is critical to effectively recognize the status of the growth of cane sugar crystals. To realize the recognition and segmentation of three cane sugar crystal states, we propose ASSugarNet, a cane sugar crystal amodal segmentation method based on deep neural networks. To better extract multi-scale semantic features, the local feature enhance module is proposed. The foreground boundary enhancement loss function is adopted, which can effectively improve the segmentation quality of sugar crystal boundary regions. For overlapping crystals, the amodal segmentation module is used to predict the contours and masks of the occluded sugar crystals, which provides accurate input for the subsequent sugar crystal growth stage judgment. To verify the network performance, we create a dataset of cane sugar crystals. Compared with FCN, PSPNet and U-Net 3 + network, the experimental results show that the proposed network structure has better segmentation for three types of crystal performance, with MIoU of 81.47% and OA of 91.00%. The AP of ASSugarNet for adhesive particle segmentation is 32.25%, and the AP50 is 52.14%.