Concerning Imbalance and Bounding Box Loss to Detect Small Targets in Remote Sensing

In general, the small target is small in scale and occupies only a small portion of a remote sensing image (RSI). Unlike large- or medium-sized objects, the small targets usually do not own complete semantic information, and the small target tends to fall into the problem of sample imbalance. The adverse factors intensify the seriousness of the challenges to build an appropriate model for small object detection (SOD). To this end, we design several effective modules and aggregate them to generate a robust SOD detector in this article. The contributions of the proposed detector are summarized as follows: 1) to effectively deal with the issue of imbalance, a dynamic sample reweighted systematic leverage function is designed to amplify the weight of hard samples; 2) we reformulate a robust loss function via merging the normalized Gaussian–Wasserstein distance with the regression loss function to remedy the shortage of complete intersection-over-union (CIoU) successfully; and 3) the two proposed functions are systematically aggregated and deployed on a selected anchor-based model to form an advanced SOD detector. The extensive experimental results demonstrate the effectiveness and efficiency of our proposed detector. Specifically, our method achieves the mAP50 values of 70.1%, 92.6%, and 95.1% on the VEDAI, NWPU VHR-10, and RSOD datasets, respectively. The results indicate that our proposed detector is superior to the recent mainstream methods.