A dual-network self-optimizing threshold method for web images

In fine-grained image recognition, the limited availability and prohibitive cost of high-quality annotated data impede the creation of large-scale, accurate datasets. To circumvent this, researchers have turned to publicly accessible web images. However, these images frequently come with noise labels, and the resulting noise instances disrupt the training process, causing a drop in recognition performance. At present, the common practice of using labels from low-loss instances for training is fraught with issues, such as a heightened risk of overfitting and compromised generalization ability. To tackle these issues, this paper proposes a new method called DST, which is designed to boost the performance of fine-grained image recognition using web images. Specifically, in the first stage, the dual-network structure is used to carry out preliminary learning of the training data, so that the model can master the basic feature expression and category discrimination ability, laying a foundation for the subsequent stage to deal with noisy labels. next, the two networks employ each other's self-optimization threshold technology. Meanwhile, they pick out more reliable samples for each other to train on. In this way, the robustness and recognition ability of the model are incrementally improved. Our approach exhibits state-of-the-art performance. It efficiently reduces the negative impact of noisy labels on model training. When compared to existing methods that utilize the same backbone network, it demonstrates significant superiority, along with robust stability and generalization ability.