Multi-scale Few-Shot Classification Model Based on Attention Mechanism

Yi Xu, Qisheng Zhu, ZhengYue Pan, Yin Liu

Abstract: Training a powerful classification model with limited examples is a key challenge for few-shot learning. Limited examples with insufficient supervised information and insufficient feature extraction cause poor classification performance. Therefore, we constructed the multi-scale features of samples based on the attention mechanism to obtain more supervised information, and at the same time, we used the Poolformer architecture as the backbone network for feature extraction, fully extracted the sample features, and defined the multi-scale less-point classification model based on the attention mechanism (AMLN). Firstly, the embedding representation of the sample is extracted through the Poolformer architecture. Secondly, the embedding representation is processed using the attention mechanism to obtain the multi-scale features of the sample. Thirdly, few-shot classification is performed on the embedding representation of each scale level through a prototype network, and the classification result of each scale level is obtained. Fourthly, we define a multi-scale synergistic loss function to make the classification results at different scale levels consistent. The loss function requires that the scale-level predicted values before attentional mechanism processing should be not only as close as possible to the true values, but also be as close as possible to the predicted values of other scale levels. The effectiveness of AMLN is proved by comparing it with existing mainstream few-shot classification models.