Computer Vision-Based Intelligent Non-Destructive Testing for Fish Meat Quality Assessment: A Hybrid Convolutional Neural Network and Support Vector Machine Approach

Abstract

Ensuring food safety has historically been one of humanity's most significant challenges, often accompanied by considerable costs, time, and potential for errors. The verification and classification of fish meat traditionally rely on destructive testing methods, such as physical and chemical tools, which present various problems and challenges. Today, with the expanding application of artificial intelligence, it is possible to classify fish meat with high speed and accuracy. Therefore, the aim of this research was to present a non-destructive intelligent model for the classification and identification of fish meat into three classes: healthy, normal, and spoiled. The dataset for modeling and analysis was collected from the Kaggle database, with each class containing 400 images. This study analyzes and implements two deep learning-based approaches for detecting fish freshness from images. The first method involves training a Convolutional Neural Network (CNN) model using the MobileNetV2 architecture with added dense layers. In the second approach, MobileNetV2 is utilized as a feature extractor, and then a Support Vector Machine (SVM) classifier is trained on the extracted features. The results indicate that both approaches achieve acceptable accuracy, but the hybrid MobileNetV2 + SVM model demonstrates significantly superior performance in classifying fish freshness categories. This study emphasizes the effectiveness of transfer learning and the advantage of using SVM for final classification in practical applications.