Incremental Learning for Fruit and Vegetable Recognition: Addressing Intra-Task Confusion and Forgetting

The task of fruit and vegetable recognition must adapt to the increasing variety resulting from hybridization. Incremental Learning (IL) is well-suited for this scenario. However, IL faces the challenge of catastrophic forgetting, where previous knowledge is gradually forgotten. Additionally, the high visual similarity between categories will lead to confusion and require advanced techniques to capture subtle distinctions. We propose General Knowledge Transfer (GKT), which freezes specific layers after the first task to preserve the shared representations learned during earlier tasks. GKT reduces the risk of catastrophic forgetting. Furthermore, we use a Gaussian Mixture Model (GMM) to save the feature distribution of each class. To reduce confusion between similar classes, we propose Similarity Contrastive (SC), which pulls old samples closer to their corresponding GMM centers and distances them from other GMMs.