Lightweight and Trustworthy Deep Learning Model for Fish Freshness Detection: An XAI-Driven Approach

Fish is a vital source of protein in Bangladesh and globally, with cultural significance in the region. As much of the population lives in urban areas, fish must be transported from rural sources, leading to potential quality loss due to poor handling. Many consumers struggle to assess freshness, creating a need for an affordable, automatic method. Traditional techniques are labor-intensive and unreliable, highlighting the demand for a more accessible solution. In this study, we propose a trust-worthy and lightweight deep-learning framework for detecting fish freshness, designed to be both practical and convenient for everyday consumers. To build an effective model, we need a comprehensive dataset that includes a wide variety of fresh and non-fresh fish. Unfortunately, no suitable dataset is publicly available, making it challenging to develop a robust model. An incomplete dataset can lead to model overfitting, where the model makes decisions based on irrelevant or incorrect features. To tackle this challenge, we propose an explainable AI-based deep learning framework that ensures the model learns proper knowledge based on the correct features. We experimented with four different CNN models—MobileNetV2, ResNet50, AlexNet, and DenseNet121—to evaluate their accuracy. Grad-CAM was used to visualize the features driving their decisions. While the models initially performed well on the Kaggle fish gill dataset, Grad-CAM revealed that their learning approach was flawed. To correct this, we guided the models to focus on the true region of interest, ensuring their decisions were based on relevant features. After implementing this new approach, all four models performed significantly better, with proper justification from Grad-CAM. Upon further analysis and comparison, we found that the MobileNetV2 model strikes the best balance between accuracy and computational efficiency. It achieved an accuracy of 99.50%, precision of 100%, recall of 98.75%, and an F1-score of 99.37%, making it well-suited for use even on low-powered mobile devices.