Convolutional Neural Network (CNN)-Based Image Analysis for Differentiating Fresh and Frozen-Thawed Vermilion Snapper (Rhomboplites aurorubens)

Abstract

Color, as an essential visual quality characteristic of fish that determines consumer preferences and purchasing decisions, serves as a quick indicator of freshness. However, determining whether fish has undergone previous freezing poses a challenge. Traditional methods such as sensory evaluation, physicochemical, biochemical, and microbiological analyses, despite being time-consuming and costly, have been employed for this purpose among various fish species. This study aims to differentiate frozen-thawed (at −20°C and −60°C) and fresh fish samples through deep learning-based image analysis. A dataset comprising frozen-thawed and fresh Vermilion snapper (Rhomboplites aurorubens) images was used to extract features with four pretrained CNN models (VGG16, ResNet50, InceptionV3, and EfficientNetB0), followed by a neural network for classification. For comparison, the same classifier was applied to traditional L∗, a∗, and b∗ color values (commonly used in conventional analyses) extracted from the images. The CNN-based models consistently outperformed the traditional approach. In the binary classification task (fresh vs. frozen), VGG16 achieved the highest accuracy of 96.8%, compared to 80.0% using color features. In the more challenging three-class classification (fresh vs. −20°C vs. −60°C), EfficientNetB0 achieved an 86.9% accuracy, outperforming the 64.0% accuracy of the color-based approach. Additionally, for left- vs. right-side classification, ResNet50 reached a 96.0% accuracy, while color-based classification achieved only 52.0%. These findings suggest that CNN-based image analysis can be a valuable tool for the seafood industry, providing a nondestructive, faster, more reliable alternative to conventional methods for assessing the freshness and freezing history of fish products.