Proposed Novel Fish Freshness Classification Using Effective Low-Cost Threshold-Based and Neural Network Models on Extracted Image Features

The quality of food has been becoming a great concern not only in Vietnam but also all over the globe. Quality of fish in terms of fish freshness is therefore highly attracted by the research and industry community. This paper proposes novel fish freshness classification models based on threshold-based and neural network-based approaches on extracted image features. These features are identified based on physiological characteristics of fish eyes at the fresh and stale statuses, including 12 Intensity Slices, Minimum Intensity, Haziness, Histogram, and Standard Deviation. The nine proposed models consisting of 4 threshold-based and 5 neural network-based models were trained on the training set composing of 49 fisheye images of the 4 Crucian carp fishes at two main groups of time points (0–5 h and 21–22 h after death) and tested on the testing set including 18 images from the fifth fish sample. The results of 8/9 models reach their 100% of accuracy on the training set and 7/9 at their 100% of accuracy on the testing set. These results confirm our four proposed feature assumptions and reveal the feasibility of the proposed models based on extracted features which are non-invasive, rapid, low cost, effective and environmental-effect minimized and consequently, highly potential for further studies and mobile application for freshness classification.