Aiming at the problems of inaccurate discriminative feature acquisition and insufficient use of training data in fine-grained image classification tasks, a fine-grained image classification algorithm combining significance and non-local modules is proposed. By clipping the significance region of four training images and stitching them into one training image, the training data could be enriched and enhanced. Moreover, the non-local module was embedded into the bottleneck module in the high-dimensional feature layer of the ResNet-50 model to connect the four salience regions of the enhanced image, which strengthened the model's attention to the global context information. On the Stanford Cars and CUB-200-2011 data sets, the accuracy of Top-1 classification was 94.01% and 85.97%, respectively. This method performs better than compared data augmentation methods and fine-grained image classification algorithms.