Classifying the freshness of large yellow croaker (Larimichthys crocea) at 12- and 24-hour intervals using computer vision technique and convolutional neural network

Abstract

To develop a rapid and non-destructive method for assessing the freshness of large yellow croaker, a computer vision technique combined with a convolutional neural network (CNN) was utilized. Sixty fish were stored on ice, and images were captured using a smartphone at intervals of 0, 12, 24, 36, 48, 72, and 96 h. A modified ResNeXt architecture was applied to automatically extract features and establish a freshness classification model. The CNN model was able to identify imperceptible visual changes, and achieved classification accuracies of 84.0 % and 72.0 % for 24- and 12 h intervals, respectively. Furthermore, potential mechanisms for the model's performance were discussed, indicating that changes in skin, eyes, and other image features contribute to the freshness classification. In summary, this method is effective for real-time, non-destructive, low-cost, and environmentally friendly fish freshness evaluation, particularly during the early stages of storage.