跨域标记牌文字检测算法研究

CROSS-DOMAIN TAG BOARD TEXT DETECTION ALGORITHM

由于受到光照、遮挡、倾斜等诸多因素的影响,基于深度学习的文字检测方法的训练集和测试集分布存在差异,导致模型在不同真实场景下的鲁棒性不足。为了提升现有模型的泛化能力,改善模型在真实场景下的抗干扰能力,尝试从网络结构角度出发,为机场标记牌文字检测任务设计具有域不变性的网络结构,在不增加计算量的前提下提升算法鲁棒性。结果表明,该结构可以有效提升模型在不同真实场景下的性能表现。

Due to the influence of illumination,occlusion,tilt and other factors,the distribution of training set and test set of text detection method based on deep learning is different,which leads to the insufficient robustness of the model in different real scenes.In order to improve the generalization ability of the existing models and the anti-jamming ability of the models in real scenes,we designed a domain invariant network structure for the text detection task of the airport tag board from the perspective of network structure to enhance the robustness of the algorithm without increasing the computation.The results show that the structure can effectively improve the performance of the model in different real scenes.