Class-Balanced Loss for Scene Text Detection

Abstract

To address class imbalance issue in scene text detection, we propose two novel loss functions, namely Class-Balanced Self Adaption Loss (CBSAL) and Class-Balanced First Power Loss (CBFPL). Specifically, CBSAL reshapes Cross Entropy (CE) loss to down-weight easy negatives and up-weight positives. However, CBSAL ignores gradient imbalance that CE gives positives and negatives different gradients. Since text detectors need to identify text and background simultaneously, positives and negatives have same importance and should possess equivalent gradients. Thus CBFPL provides equal but opposite gradients for positives and negatives to eliminate this gradient imbalance. Then, CBFPL abandons easy negatives and makes their gradients zero to handle class imbalance. Both CBSAL and CBFPL can focus training on positives and hard negatives. Experimental results show that on the basis of CBSAL and CBFPL, the efficient and accurate scene text detector (EAST) can achieve higher F-score on ICDAR2015, MSRA-TD500 and CASIA-10K datasets.