Aiming at the image classification problem in small-sample learning, we study the classification algorithm based on the image deformation network, and propose the improvement strategy of modifying the way of selecting auxiliary graphics and modifying the value of image fusion weights. The strategy combines the relational network with the Euclidean distance and includes the RGB channel parameters in the training range as well. On this basis, the dynamic adaptive fusion strategy is further introduced to avoid the target region being segmented too finely. The experimental results show that the image deformation meta-network achieves the best performance in both 1-shot and 5-shot classification tasks, with accuracies of 59.17% and 74.66%, respectively. The improved relational network-deformation meta-network shows improved performance by achieving 59.92% and 75.23% accuracy in 1-shot and 5-shot tasks, respectively. Further the network with dynamic adaptive fusion strategy achieves an accuracy of 60.15% ± 0.28 on the 5-way 1-shot small sample image classification task, which is significantly better than the other strategies. The experimental results show that the improvement of small-sample image classification algorithm based on image deformation meta-network is effective and can significantly improve the classification accuracy, especially the dynamic adaptive fusion strategy, which plays a positive role in improving the classification accuracy.
Lei et al. (Sat,) studied this question.