Entropy-Based Sampling Strategy for Long-Tail Target Detection of SAR Images

SAR images reflect the scattering characteristics of targets, whereas the quantity and the characteristics vary greatly from categories of targets, which causes the extreme inter-class imbalance problem. Most target detection methods for SAR images pay more attention to the algorithms but neglect the imbalance problem caused by the dataset itself. Therefore, an entropy-based sampling strategy with logarithmic smoothing is proposed to solve the extreme long-tail class imbalance problem for SAR detection tasks. Local image entropy is introduced to evaluate the information quantity or difficulty of a SAR image, and the relationship between them has been discussed elaborately. Logarithmic smoothing is utilized to avoid excessive scores due to ignoring the marginal effect. The results on the MSAR dataset also illustrate the best performance of the proposed method, which proves that the proposed method considered the quantity and the entropy simultaneously will get better performance.