Combining spatial attention and cross-layer bilinear pooling for fine-grained image classification

To address the problems of large intra-class difference and small inter-class-class difference in fine-grained images and the difficulty of obtaining effective feature representations, this paper proposes a method combining spatial attention and cross-layer bilinear pooling for fine- grained image classification, which can learn a more powerful fine-grained feature representation. In the paper, extract the underlying image features through ResNet50 network and inspired by the multi-scale detection algorithm of small-size objects, a spatial attention mechanism is used to achieve adaptive adjustment of the spatial weights of each layer features in fusion, which promotes more effective fusion of higher-level features and lower-level features to complement each other. In addition, cross-layer bilinear pooling is proposed in the paper to integrate the features of each layer after two- by-two interaction in order to explore the correlation of features between layers. Finally, by evaluating on three publicly available datasets, the experimental results indicate that the proposed method is superior to the existing classical methods in terms of classification.