引入注意力机制的自然场景文本检测算法研究

NATURAL SCENE TEXT DETECTION ALGORITHM WITH ATTENTION MECHANISM

随着深度学习、神经网络的兴起与发展,对于图像中的目标检测已经取得了巨大的进展。但是自然场景下的文本信息具有多样的形式和复杂的特点,通用的目标检测算法无法取得理想的效果,因此自然场景下的文本检测在计算机视觉以及机器学习领域仍然是一项具有挑战性的问题和未来的热点研究方向。根据当前学术界针对自然场景下的文本检测问题所提出的算法和思路,在EAST算法的主干网络PVANet的基础上通过引入注意力机制模块,使得提取文本目标特征时更加关注有用信息和抑制无用信息,从而有效改善原算法在预测长文本方向信息时视野不足的问题。实验结果显示,该方法在没有损失检测效率的同时提高了原算法的检测精度,并在一定程度上优于当前针对自然场景下的文本检测算法。

With the rise and development of deep learning and neural network, object detection in images has made great progress. However, text information in natural scenes has various forms and complex characteristics, so universal object detection algorithms cannot achieve ideal results. Therefore, text detection in natural scenes is still a challenging problem and a hot research direction in the field of computer vision and machine learning. This paper studied the current academic algorithms and ideas for text detection in natural scenes. Based on the network-PVANet of EAST algorithm, we introduced the attention mechanism module to make the extraction of text target features pay more attention to useful information and suppress useless information, so as to effectively improve the problem of insufficient field of vision of EAST in predicting long text direction information. The experimental results show that compared with EAST, this method improves the detection accuracy without losing the detection efficiency, and to some extent it is superior to the current text detection algorithm for natural scenes.