Cell Image Incremental Classification with Memory-Augmented Convolutional Neural Networks
Qi Jin, Kuo-Kun Tseng · 2021 IEEE 4th International Conference on Automation, Electronics and Electrical Engineering (AUTEEE) · 2021
The traditional manual analysis method in the field of cell image recognition is time-consuming and laborious, and Convolutional Neural Networks (CNNs) can already handle this type of work well. However, subject to the catastrophic forgetting problem, current CNNs cannot adapt to incremental learning scenarios, which is quite common in cell image classification. In this paper, we propose a new method to handle this problem. Our method mainly consists of an autoencoder and a fixed-capacity memory bank. On one hand, the autoencoder adopts the popular CNN structure, the encoder encodes the training image as feature vectors which will be saved into the memory bank during training. The decoder can generate old categories knowledge based on the memory bank to against catastrophic forgetting when learning new categories of cells. On the other hand, the memory bank saves the feature vectors of each learned category and can be a nearest-neighbor classifier to better adapt to the incremental scenario. We design a specific updated algorithm for the memory bank to make full use of its limited capacity. We conduct comparative experiments with other state-of-the-art methods on cell image classification datasets and the results demonstrate the superiority of our method.