In factories that manufacture round bars, an accurate count of the number of bars is required before shipment. Manual counting is time-consuming and labor-intensive, and the larger the number of bars, the less accurate is the count. Therefore, we automated the counting process using semantic segmentation. The method introduces a novel idea to reduce the effort of annotating the training images and generating the ground truth. Specifically, points are manually annotated on each end face of a round bar, and ground truth is automatically generated from the points. The segmentation model was trained using the generated ground truth to extract each end face from the target image. The round bars were counted by applying labeling after removing salt-and-pepper noise. To confirm the effectiveness of our method, an image dataset was created. In experiments, the performance of DeepLabV3+ and Unet++ with conflicting features were compared. The results showed that Unet++, which considers local information, performed better.
Sato et al. (Thu,) studied this question.