Due to subtle visual differences among fine-grained subcategories only locate at local regions, how to locate and extract differentiated local features has become the primary contradiction to be solved in this field. In the early stage, many works based on weak supervision are inclined to focus only on the most discriminative feature while ignoring other features that play an equally important role in image recognition. To tackle this problem, this paper proposes an adversarial complementary strategy-based method that can urge the network to find a complementary region by removing the most meaningful parts in the image, which can extract more information with significant features to promote the precision of fine-grained image classification. At the same time, similarity loss was utilized to determine the similarity between the two regions to avoid obtaining similar features as much as possible. In addition, Bilinear Attention Pooling (BAP) is introduced to extract sequential part features, which can effectively merge several different levels of features together. Experiments show that our method achieves the competitive results in various challenging datasets, such as CUB-200-2011, FGVC-Aircraft and Stanford Car.