Multi-stream Fusion for Class Incremental Learning in Pill Image Classification

AbstractClassifying pill categories from real-world images is crucial for various smart healthcare applications. Although existing approaches in image classification might achieve a good performance on fixed categories, they fail to handle novel instances of that are frequently presented the learning algorithm. To this end, trivial solution train model with classes. However, may result phenomenon known as catastrophic forgetting, which system forgets what it learned previous In paper, we address challenge by introducing class incremental (CIL) ability traditional systems. Specifically, propose multi-stream intermediate fusion framework enabling incorporation an additional guidance information stream best matches domain problem into state-of-the-art CIL methods. From framework, consider color-specific and devise approach, namely “Color Guidance Multi-stream fusion” (CG-IMIF) solving task. We conduct comprehensive experiments dataset, VAIPE-PCIL, find CG-IMIF consistently outperforms several methods large margin different task settings. Our code, data, trained available at https://github.com/vinuni-vishc/CG-IMIF.