Abstract:

A deep neural network based on fused loss function is proposed to solve the shortcomings of natural scene text detection based on traditional deep neural network, such as poor detection, inaccurate detection of text bounding boxes and so on. The Balanced loss is combined with the traditional deep neural network by using weighted method to increase the loss value of pixels at the border of the text boxes and those difficult to detect. Thereby, the optimization direction of the model is constrained, and the ability of the model to learn complex features has improved. The experimental results show that the proposed method effectively improves the detection accuracy of the network in the natural scene text detection.

Key words: neural network, text detection, loss function, deep learning

摘要：

针对基于传统深度神经网络的自然场景文本检测存在检测效果较差、文本边界框检测不准确等缺陷，提出基于损失函数融合的深度神经网络。将损失函数Balanced loss，利用加权的方法与传统深度神经网络进行融合，用于提高文本框边界区域及图像中难检测像素点的损失值，从而约束模型的优化方向，提升模型学习复杂特征的能力。实验结果表明，在自然场景文本检测中所提出方法有效提高了网络的检测准确性。

关键词: 神经网络, 文本检测, 损失函数, 深度学习