SaDiTect: Dirt Detection in Salt Using YOLOv3

Salt is essential for maintaining people's life. The presence of undesirable pieces such as dirt contributes to the overall quality of salt provided to customers. People's naked eyes seldom distinguish between salt and dirt; as a result, it would take time and effort to separate salt and dirt. Unseen dirt can also add to the total weight of manufacturing fine salt, and removing this dirt manually can be time-consuming. Artificial Intelligence (AI) algorithms employ object categorization systems to recognize certain items in a class as the topic of study. The systems aggregate things in pictures into categories where objects having similar qualities are grouped along, while others are ignored until specifically configured. The objective of the study is to develop a detection system for unwanted dirt that is mixed in salt. This system can be embedded into a dirt removal system for the salt manufacturing process. The study gathered 500 images of salt as the dataset and divided it into two (2) parts: training is set to 70% and for testing is 30%. creating the model using the dataset that has been gathered together with the Yolov3 pre-trained model for object detection was used in creating the model for the dirt detection system and the training has 50 epochs. The researchers conducted testing using ten (10) photos of dirt to assess the system's accuracy, and they reached a 70 percent accuracy. This study also evaluated the system by importing video clips to be detected and the system easily detected most of the dirt. This demonstrates that the system is trustworthy and effective in detecting dirt in the salt. This system can be improved in terms of accuracy by adding techniques like data augmentation, transfer learning, and model selection.