Fish grades identification system with ensemble-based key feature learning

Abstract

Indonesia has already contacted the maritime nations due to its 5.8 million km2 of coastline. Consequently, fish products are among the most important commodities. Moreover, fish grading is a crucial step in the process of exporting fisheries products. Currently, in Indonesia, the process itself is manually inspected by an expert. In addition, this paper proposes to assist the industry by suggesting a method for grading fish. This method involves combining two essential fish parts with different resolutions: the high-level feature (the body) and the low-level feature (the eye) serve as defining characteristics. These two main parts are accurately localized using a deep learning-based object detection model, specifically YOLOv7, and extracted with an efficient and adaptive learned classification model, namely EfficientnetV2S. In the final stage, the two extracted features are combined and learned with Dense Layers to generate three distinct fish grades. Based on the experimental results, the proposed work achieved an accuracy, F1 Score, and recall of 96.88%, 97%, and 97%, respectively. The proposed model outperformed the baseline model, which relies solely on deep learning-based classification, by a significant margin.