Iced pomfret freshness evaluation method based on improved VGG-19 convolutional neural networks

Abstract: The quality and safety of iced aquatic products have been the key steps to improve the benefits in the aquatic industry. However, there are three challenges in the traditional evaluation of the freshness of aquatic products, including the complicated operations, samples destruction, and low efficiency. It is highly urgent to find an effective way for the aquatic products in a cold chain system. Taking the iced pomfret as a research object, this study aims to propose a novel computer vision for the evaluation of the freshness using an improved VGG-19. The dataset collection and algorithm design were also utilized as follows: 1) The image dataset of the freshness grade was established to integrate the environmental factors and the Total Volatile Basic Nitrogen (TVB-N) in the cold chain system. Specifically, the image data was collected, according to the physicochemical index TVB-N of freshness at a constant temperature of 0 ℃, with days as the unit of time. As such, a total of 2 387 image samples were collected after 30 days. The freshness level of pomfret was then divided into three categories, such as the first-class, qualified, and unqualified, according to the standard regulations of SC/T 3103-2010 "Fresh and Frozen Pomfret". 2) Deep learning has developed rapidly in recent years. Among them, Convolutional Neural Networks (CNN) can automatically learn a large number of samples and extract features of interest, which are widely used in various image classification tasks. Four CNN models were selected for the classification of the freshness, including AlexNet, VGG-16 (visual geometry group 16), VGG-19 (visual geometry group 19), and ResNet-50 (Residual Network-5). The weights and biases of the pre-training model were fine-tuned to improve the generalization of the model. After that, the VGG-19 with the transfer learning was selected as the main model of the classification. More importantly, the dataset of freshness classification was designed for the three-classification task. The fully connected layers were modified to save time and space costs for the higher efficiency and accuracy of the network. The reason was that the three-layer fully connected layer of traditional VGG-19 presented a complex structure with too many parameters. After modification, the improved network was utilized to rapidly and accurately classify the features. After comparison, the 4 096, 4 096, and 1 000 three-layer fully connected structure of VGG-19 was improved to the 128, 3 two-layer fully connected structure. The experimental results show that the improved model achieved a 99.79% accuracy of classification with less time and space, where the parameters of the fully connected layer were reduced by 97%, and the space was reduced by 443.9 MB. In addition, a class activation mapping was introduced to visualize the freshness grading, thereby understanding the grade judgment of freshness in the model. Subsequently, 18 pomfret pictures were selected with the different freshness levels, regions, and orientations. A confusion assessment method (CAM) was also applied to generate a visual heat map for subsequent experimental analysis. Consequently, the improved CNN model can be expected to accurately predict the freshness level with a specific basis for the judgment. Among them, the most effective information for the freshness evaluation was the abdominal characteristics of pomfrets, no matter where the pomfret was placed. This improved model was also suitable for the various iced aquatic products, particularly when it is difficult for the naked eye to determine the freshness level in the aquatic industry.