Deep Learning for Fish Freshness Detection: An XAI Oriented Knowledge Validation Based Approach

Fish is a vital source of protein and essential nutrients both in Bangladesh and globally. Consuming fresh fish preserves its nutritional benefits and supports good health, whereas non-fresh fish can lose key nutrients and may become harmful depending on the level of spoilage. Typically, fish are sourced from rivers and seas or are farmed in ponds and lakes. However, many people live far from these fish-producing regions, necessitating extended transportation and storage, which increases the risk of quality degradation. In urban areas, consumers often struggle to accurately assess fish freshness, sometimes being misled into purchasing spoiled fish. This situation highlights the need for an affordable and automated method to determine fish freshness reliably. To address this issue, we propose a lightweight deep-learning framework designed for deployment on low-configuration smartphone devices, ensuring accessibility and practicality. Developing a robust deep-learning model typically requires a high-quality dataset; however, no standard dataset currently exists for fish freshness detection. To overcome this limitation, we introduce an explainable AI-driven framework that leverages Grad-CAM-guided feature activation during model training. This approach enhances model reliability and interpretability, ensuring trustworthy performance even when using limited or non-standard datasets. By ensuring the model focuses on relevant regions of interest, our approach enhances both reliability and interpretability, even when trained on a small dataset. Experimental results validate the effectiveness of our framework, achieving over 99% accuracy. These results underscore the model’s robustness and practical applicability, offering a promising solution for ensuring fish quality in the supply chain.