Improved fine-grained image classification in few-shot learning based on channel-spatial attention and grouped bilinear convolution

In the context of the complexities of fine-grained image classification intertwined with the constraints of few-shot learning, this paper focuses on overcoming the challenges posed by subtle inter-class differences. To enhance the model's capability to recognize key visual patterns, such as eyes and beaks, this research ingeniously integrates spatial and channel attention mechanisms along with grouped bilinear convolution techniques to adapt to the few-shot learning environment. Specifically, a novel neural network architecture is designed that integrates channel and spatial information, and interactively applies these two types of information to collaboratively optimize the weights of channel and spatial attention. Additionally, to further explore the complex dependencies among features, a grouped bilinear convolution strategy is introduced. This algorithm divides the weighted feature maps into multiple independent groups, where bilinear operations are performed within each group. This strategy captures higher-order feature interactions while reducing network parameters. Comprehensive experiments conducted on three fine-grained benchmark datasets for two few-shot tasks demonstrate the superiority of our algorithm in handling fine-grained features. Notably, in the experiments on the Stanford Cars dataset, a classification accuracy of 95.42% was achieved, confirming its effectiveness and applicability in few-shot learning scenarios. Codes are available at: https://github.com/204503zzw/atb.