基于部分加权损失函数的RefineDet

针对目标检测网络单阶改进目标检测器（RefineDet）对类间不平衡数据集中小样本类别检测性能差的问题，提出一种部分加权损失函数SWLoss。首先，以每个训练批量中不同类别样本数量的倒数作为启发式的类间样本平衡因子，对分类损失中的不同类别进行加权，从而提高对小样本类别学习的关注程度；然后引入多任务平衡因子对分类损失和回归损失进行加权，缩小两个任务学习速率的差异；最后，在目标类别样本数量存在大幅差异的Pascal VOC 2007数据集和点阵字符数据集上进行实验。结果表明，与原始RefineDet相比，基于SWLoss的RefineDet明显提高了小样本类别的检测精度，它在两个数据集上的平均精度均值（mAP）分别提高了1.01、9.86个百分点；与基于损失平衡函数和加权成对损失的RefineDet相比，基于SWLoss的RefineDet在两个数据集上的mAP分别提高了0.68、4.73和0.49、1.48个百分点。

RefineDet based on subsection weighted loss function

Concerning the poor performance of the Single-Shot Refinement Neural Network for Object Detection (RefineDet) of the object detection network when detecting small sample classes in inter-class imbalanced datasets, a Subsection Weighted Loss (SWLoss) function was proposed. Firstly, the inverse of the number of samples from different classes in each training batch was used as the heuristic inter-class sample balance factor to weight the different classes in the classification loss, thus strengthening the concern on the small sample class learning. After that, a multi-task balancing factor was introduced to weight classification loss and regression loss to reduce the difference between the learning rates of two tasks. At last, experiments were conducted on Pascal VOC2007 dataset and dot-matrix character dataset with large differences in the number of target class samples. The results demonstrate that compared to the original RefineDet, the SWLoss-based RefineDet clearly improves the detection precision of small sample classes, and has the mean Average Precision (mAP) on the two datasets increased by 1.01 and 9.86 percentage points, respectively; and compared to the RefineDet based on loss balance function and weighted pairwise loss, the SWLoss-based RefineDet has the mAP on the two datasets increased by 0.68, 4.73 and 0.49, 1.48 percentage points, respectively.