基于部分加权损失函数的SSD仪表字符检测方法研究

Research on SSD instrument character detection method based on partial weighted loss function

针对变电站中数字仪表字符受环境、类间数量不平衡导致的识别准确率底的问题,文章提出了一种基于部分加权损失函数的SSD字符检测方法。文章先通过对SSD的分类损失函数引入类间样本平衡因子,提升少样本的仪表字符类别的关注度;再结合多任务平衡因子,调节分类损失和回归损失之间对权重更新速率差异。实验以存在类间样本数量不平衡的仪表字符为对象,与SSD进行对比,基于部分加权损失函数的SSD对仪表字符检测,mAP-50提升了3.23个百分点。

Aiming at the low recognition accuracy of digital instrument characters in substation caused by the unbalanced number of environment and classes, this paper proposes a SSD character detection method based on partial weighted loss function. Firstly, this paper introduces the sample balance factor between classes into the classification loss function of SSD to improve the attention of instrument character categories with few samples;Combined with multi task balance factor, the difference of weight update rate between classification loss and regression loss is adjusted. The experiment took meter characters with unbalanced sample number between classes as the object, and compared with SSD. The detection of meter characters based on SSD with partial weighted loss function improved the map-50 by 3.23 percentage points.