CNN-Based Real-Time Fish Species Identification and Freshness Assessment for Food Safety

Abstract

Fish is one of the most wanted food in coastal areas, but fish freshness identification and fish species identification is very challenge. The accuracy of fish identification and freshness is of significant importance for food safety, traceability, and automated processing of fish. In this paper, we have introduced FishSense, a lightweight multi-output CNN model based on MobileNetV2 that is capable of carrying out simultaneous fish identification and binary freshness prediction from a single image. The model is trained on a dataset of 12,480 images containing eight different fish species and two different freshness states. The multi-output model has been able to achieve an accuracy of 96.4% for fish identification and a 94.2% F1 score for freshness prediction. The performance of the model is enhanced compared to its single-output counterparts by an increase of 2.8% and 3.1%, respectively. The model has 3.5 million parameters and is able to achieve 32 FPS on an NVIDIA Jetson Nano, which is 31 ms per image. This shows that the model is appropriate for real-time application.