Robust small-sample classification via RN50STN architecture with adaptive example reweighting

Convolutional Neural Networks (CNNs) have shown exceptional performance across various domains; however, their deep architectures require extensive datasets for effective training. In practical situations, gathering a large number of training examples can be extremely challenging, and the model is at risk of overfitting due to the limited and potentially biased nature of the available data. To address this challenge, a novel approach using a deep migration hybrid model has been introduced to tackle the issues posed by small-sample datasets. Initially, transfer learning is utilized to convey the knowledge acquired by a pretrained network from the ImageNet database to a limited dataset. Subsequently, the fully connected layers and adaptive average pooling layers of the pretrained network are augmented by incorporating a spatial transformer network (STN) for further training. Simultaneously, an online reweighting algorithm for examples is employed to reduce sample bias and enhance the robustness of the model’s classification capabilities. The proposed algorithm notably improves the classification accuracy of small-sample, imbalanced datasets, as demonstrated by comparing various evaluation metrics before and after its implementation. This confirms that our method is resilient to class imbalance and sample corruption.