The Imbalanced Recognition of Space Micro-Motion Targets Based on GRB Loss

Limited by the quantity of space micro-motion targets launches and observation conditions, datasets for recognizing space micro-motion target usually is class imbalanced. Under such circumstances, the recognition performance of the available methods degrades rapidly, and in some cases, even complete failure. To tackle this issue, this paper proposes a gradient-guided class re-balancing loss (GRB Loss) for the imbalanced datasets of space micro-motion targets recognition. Particularly, in the proposed GRB Loss, we design an adaptive weighting function, which use the accumulated positive and negative gradient ratio of each classification node as input, to assign the weights to the positive and negative gradient, so as to improve the imbalance of the positive and negative gradient ratio in each classification node. Experimental results show that the proposed method enhances recognition performance on imbalanced datasets and remains robust across varying imbalance ratios.