ConvNeXt-ECA: An Effective Encoder Network for Few-Shot Learning

After having introduced a comprehensive general solution framework for few-shot learning (FSL) classification problems, we provide details of the data augmentation schemes and the learning techniques used, which are critical in minimizing the risk of overfitting, due to the fact that only a few labeled samples are available to fine-tune a robust encoder. Once such a general solution framework is in place, we obtain a convolutional neural network (CNN)-based encoder, which is evolved along a number of architectural brick axes from a baseline ResNet-50 model, while in each step of this evolution process, the adoption or not of the temporary incremental model will be assessed through this general solution framework. This is called a ConvNeXt encoder. We then insert an efficient channel attention (ECA) module to more effectively process the features extracted in the depthwise convolutional layer in the ConvNeXt encoder, with a minimal increase in the number of parameters used, and obtain a novel encoder: ConvNeXt-ECA. We apply the ConvNeXt-ECA encoder in the general solution framework on 5 popular FSL benchmark datasets and found that our ConvNeXt-ECA encoder achieves the best average few-shot classification accuracy when compared with other state-of-the-art results.