Underground Target Recognition Based on Cross-domain Knowledge Transfer

In the field of underground target recognition, the small sample problem limits the training effectiveness and generalization ability of deep learning models. Although generative adversarial networks can augment datasets, the discrepancy between generated images and real images tends to exacerbate overfitting. To address this issue, a cross-domain knowledge transfer model is proposed in this paper. Based on YOLOv8, the proposed model first extracts features from generated images (source domain) and then effectively transfers knowledge from the source domain to real images (target domain) through knowledge distillation techniques. Furthermore, an optimized transfer process is achieved by incorporating an improved loss function that combines masked generative distillation and Logit loss. Experimental results demonstrate that under small sample conditions, the proposed model enhances recognition accuracy from 80.60% to 92.62% and improves the mean average precision at an intersection over union of 0.5 from 0.864 to 0.908. This not only effectively mitigates the small sample problem but also significantly boosts the recognition accuracy and generalization ability of the model.