A Fast Instance Segmentation Technique for Log End Faces Based on Metric Learning

The diameter of the logs on a vehicle is critical part logistics and transportation logs. However, manual size-checking method inefficient affects efficiency log transportation. example segmentation methods can generate masks for each end face, which helps automate check gauge improve efficiency. model uses rectangle detection to identify face then traverses rectangular boxes mask extraction. traversal increases time consumption lacks separate handling overlapping areas between boxes, causes decline in extraction accuracy. To address above problems, we propose fast instance further improving accuracy log-checking diameter. convolutional neural network extract image, frame prediction embed vector image from input image. used region, generates an enveloping log, turn divides region into instances. For regions metric learning paradigm increase embedding distance pixels located different decrease same finally are instantiated according pixel vectors. This avoids repeated calls contour algorithm box enables fine delineation region. verify proposed algorithm, working pile photographed scenes using smartphone obtain recognition database divide training set, validation test set 8:1:1. Secondly, masks, ruler determined by edge-fitting combined with ruler. Finally, practicality evaluated calculating check-rule error, running speed, error wood volume calculation under national standards. has 91.2% 50.2 FPS respectively, faster more accurate than mainstream model. relative −4.62% −4.25%, −5.02%, −6.32%, −5.73% measurement Chinese, Russian, American, Japanese raw standards, respectively. Among them, calculated timber our standard smallest, indicates that this paper suitable application domestic production operations.