Stream Image Classification Using Class-wise Incremental Learning and Pre-trained Convolution Neural Network

Training deep learning architectures for computer vision tasks has required resource-intensive environments, although decreasing costs of memory and processing units have made this more feasible. However, the volume and velocity of incoming data in the context of big data make it impractical to retrain a complex convolutional neural network (CNN) on the entire dataset every time new data is received, especially for real-time processing. As the dataset grows continuously, the learning algorithm slows down and may even suffer performance degradation. To address this issue, the trained model must be updated incrementally with only new data points. This approach, however, often leads to catastrophic forgetting, where the model loses previously learned knowledge. This paper proposes a novel method to partially train the neural network using a subset of samples and fine-tune its accuracy by employing the modified version of Streaming Chunk Incremental Learning (SCIL), called CNN-SCIL. We demonstrate that this method, CNN-SCIL, not only improves accuracy but also significantly accelerates the learning process compared to fine-tuning the CNN model.