DCSGAN: A Few-shot Classification Method Based on Multi-scale Attention GAN

Generative adversarial networks (GAN) based methds have been widely used to solve the problem of scarcity of samples for few-shot classification in image recognition and natural language processing. However, when the generator can fool the discriminator by generating single samples, it will no longer generate multi-type samples, termed pattern collapse. This results in absence of diversity and poor generalization of new samples in few-shot classification. Thus, this paper proposes a few-shot automatic recognition model based on multi-scale attention mechanism GAN (DCSGAN). This model first uses multi-scale attention mechanism based GAN to extract multi-granularity feature and synthesize data with high morphological similarity. Secondly, the orthogonal vector strategy is used to dynamically adjust the weight of multiple generators to learn complementary feature, further enriching the sample diversity. Finally, convolutional neural network (CNN) is utilized to better extract the characteristics of various types of samples, enhancing the accuracy of few-shot classification. The classification accuracy of DCSGAN on MiniImageNet and TieredImageNet datasets are 0.87 and 0.92, respectively. It shows that DCSGAN-CNN performs better in few-shot classification.