Comparing Halcon and Detectron2 for detecting impurities in pig 1 intestines used as natural sausage casings

Pig intestines used as natural sausage casings are currently cleaned and then sent from Europe to 12 China for manual quality control and grading. This process can be made greener by automation that 13 removes the need for transport. A suitable machine now exists, able to check the casings for leaks 14 and to grade and cut them to standard lengths. The one remaining quality control process is 15 checking for impurities left by the cleaning process. A sophisticated vision system and a deep 16 learning system is needed. 17
After preliminary lighting tests, images of cleaned pig intestines destined for sausage casings were 18 examined manually for impurities. Pixels depicting impurities were labelled and mask impurity 19 images produced as "ground truth". Two deep learning methods were applied in order to predict the 20 areas of impurities in these images: Halcon with SOLOv2 semantic segmentation and Detectron2 21 with Mask-R-CNN instance segmentation. Despite the over-abundance of background pixels, both 22 algorithms learned the segmentation. Detectron2 was more accurate but Halcon found more of the 23 impurities, which was attributed to the difference between the segmentation types: semantic vs 24 instance. Since the aim of this work is to produce clean intestines for use in the food industry, false 25 positives are more acceptable than false negatives, so Halcon was chosen.