Refresh Learning for Dynamically Changing Fresh Produce Classification in Continual Learning Scenario

Precise and prompt product identification is vital for efficient inventory management in the retail grocery industry. While deep learning models have significantly improved in classification accuracy, dynamically changing environments, seasonal variations, and new products pose significant challenges. These factors create new data on which pre-trained models are not adequately trained, leading to a substantial drop in accuracy. Continual learning has emerged as a solution, allowing models to adapt to new data without forgetting previously learned information. However, this solution still suffers from catastrophic forgetting, especially when dealing with large datasets. This paper aims to explore strategies to mitigate catastrophic forgetting and develop a robust model for fruit and vegetable classification in the retail sector. The models used are two memory-based models (DER++ and ER) and two regularization-based models (EWC and EWC + CPR). The result indicates that memory-based models outperform regularization-based models, with Dark Experience Replay++ (DER++) showing the highest accuracy in both classincremental and task-incremental settings. Memory-based models are less affected by catastrophic forgetting as they have direct access to past data, allowing them to have better knowledge preservation. Furthermore, the study also reveals that hybrid approaches (e.g., EWC + CPR) can further enhance retention accuracy.