A fine-grained image classification method based on multiscale feature fusion is proposed. By using feature pyramid structure, the scales of different levels of features are transformed, and the information fusion is then carried out. After that, the first three regions with the most detailed features are screened out, combining with the global feature of the image to determine the subclass category of the image. The experiments are conducted on the open fine-grained data sets CUB-200-2011 and Stanford Dogs, and the classification accuracy is 85.7% and 83.5%, respectively. Experimental results show that the method has certain advantages for fine object classification.