计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (1): 204-208.DOI: 10.3778/j.issn.1002-8331.2007-0436

• 模式识别与人工智能 • 上一篇    下一篇

面向自然场景文本检测的改进NMS算法

杨有为,周刚   

  1. 新疆大学 信息科学与工程学院,乌鲁木齐 830046
  • 出版日期:2022-01-01 发布日期:2022-01-06

Improved NMS Algorithm for Text Detection in Natural Scenes

YANG Youwei, ZHOU Gang   

  1. School of Information Science and Engineering, Xinjiang University, Urumqi 830046, China
  • Online:2022-01-01 Published:2022-01-06

摘要: 近些年来,卷积神经网络算法在自然场景文本检测效果上较传统算法已经有了很大提升,但如何有效处理神经网络输出层候选框仍然值得研究。非极大值抑制算法(non-maximum suppression,NMS)通过选择最高置信度候选框作为检测结果,往往容易对较长文本以及混叠文本区域检测失效。考虑到该问题,可以将候选框集合进行排序滤波与融合计算,得到更准确的候选框,有效减少上述检测失效的情况。这种方法,可以直接嵌入原有方法中,而不需要改变网络结构或者增加任何训练量。通过在公开数据集上进行实验,对比其他方法,该方法有较大优势。

关键词: 自然场景文本检测, 卷积神经网络, 非极大值抑制, 排序滤波, 融合计算

Abstract: In recent years, the convolutional neural network algorithm has greatly improved the natural scene text detection effect compared with the traditional algorithm, but how to effectively deal with the neural network output layer candidate box is still worthy of study. The traditional non-maximum suppression(NMS) algorithm selects the highest confidence candidate box as the detection result, and it is often easy to fail to detect longer text and aliased text regions. Considering this problem, the set of candidate box can be sorted, filtered, and fused to obtain a more precise candidatebox, which effectively reduces the detection failure. This method can be directly embedded in the original method without changing the network structure or adding any training amount. By conducting experiments on public datasets and comparing with other methods, this method has great advantages.

Key words: natural scene text detection, convolutional neural network(CNN), non-maximum suppression(NMS), order filtering, fusion computing