Dried fish type identification using machine learning techniques: A case study on three fish species in Sri Lanka

Abstract

Dried fish is one of the major and traditional dishes of Sri Lankans and is also one of the major animal protein sources with low cholesterol for humans. Despite, the benefits and the popularity, people are facing challenges in identifying certain varieties of dried fish which look the same such as Rastrelliger kanagurta (Kumbala), Goldstriped sardinella (Salaya) and Sardina pilchardus (Keeri). When the above three types of dried fish are mixed, the ability to differentiate one from another becomes challenging and benefits between the three varieties is a concern. Consumers have been misled by some sellers who manipulate the challenge of identification. Therefore, the aim of this research is to identify the correct type of dried fish even if those fish are mixed together with other types. To address the above issue, a machine learning based solution was developed along with image processing using more than 3500 images. The developed model uses four features of dried fish, the head, trunk, tail, and the entire body. Using YOLOv8 model, first we attempted to identify the number of objects in the user captured image and we store those identified objects temporarily within the model. Subsequently a trained VGG16 model is used for the classification of dried fish types based on the previously identified objects. The model achieved more than 90% overall accuracy in identifying the correct dried fish type. In conclusion, the developed model can be used to effectively identify the type of dried fish even in situations where they are mixed with other dried fish types. The model will be further developed as a mobile application in the future for the betterment of dried fish consumers as well as sellers.