基于分割的自然場景下文本檢測方法與應用
2021年電子技術應用第2期
陳小順，王良君
江蘇大學 計算機科學與通信工程學院，江蘇 鎮江212013
摘要： 自然場景文本檢測識別在智能設備中應用廣泛，而對文本識別的第一步則是對文本進行精確的定位檢測。對于現有像素分割方法PixelLink中存在的彎曲文本定位包含過多背景信息、檢測圖像后處理不足兩個主要問題提出改進。引入特征通道注意力機制，關注生成特征圖中特征通道間的權重關系，提升檢測方法的魯棒性。接著改變公開數據集標注形式，將坐標點表示為一串帶有方向的序列形式，在LSTM模型中進行多邊形框的學習與框定。最后在公開數據集和自建數據集上進行文本檢測測試。實驗表明，改進的檢測方法在各數據集中表現優于原方法，與當前領先方法精度相近，能夠在各個環境中完成對文本的檢測功能。
中圖分類號： TN911.73；TP391.4
文獻標識碼： A
DOI：10.16157/j.issn.0258-7998.200316
中文引用格式： 陳小順，王良君. 基于分割的自然場景下文本檢測方法與應用[J].電子技術應用，2021，47(2)：54-57.
英文引用格式： Chen Xiaoshun，Wang Liangjun. Text detection and application in natural scene based on segmentation[J]. Application of Electronic Technique，2021，47(2)：54-57.
文獻標識碼： A
DOI：10.16157/j.issn.0258-7998.200316
中文引用格式： 陳小順，王良君. 基于分割的自然場景下文本檢測方法與應用[J].電子技術應用，2021，47(2)：54-57.
英文引用格式： Chen Xiaoshun，Wang Liangjun. Text detection and application in natural scene based on segmentation[J]. Application of Electronic Technique，2021，47(2)：54-57.
Text detection and application in natural scene based on segmentation
Chen Xiaoshun，Wang Liangjun
School of Computer Science and Telecommunication Engineering, Jiangsu University，Zhenjiang 212013，China
Abstract： Text recognition in nature scene is currently applied in various intelligence equipment. The first step of text recognition is to precisely locate the text. In the Pixel Link text location methods, there are mainly two problems: too much background information is incorporated in the text region, and the test accuracy is insufficient. Aiming at these issues, an improved text location method was proposed to precisely locate the text in the natural scene. At first, an attention mechanism was incorporated into the original network. By focusing on the weight relationship between feature channels in the generated feature map, one can improve the weight coefficient of effective feature channels, and suppress the weight of inefficient or invalid feature channels. In the second, by changing the form of data set annotation, the coordinate points can be expressed as a series of sequence forms, so that the text lines can be framed adaptively in the LSTM model. At last, the located object is rotated according to the angle between a pair of vertexes in the polygon frame, and is subsequently fed to the text recognition interface to obtain the final character. Finally, the text detection test is carried out on the open data set and self-built data set. The experimental results show that the improved detection method is superior to the original method on different dataset, and the accuracy is similar to the current leading method.
Key words : pixel segmentation；attention mechanism；LSTM；natural scene text detection
0 引言
視覺圖像是人們獲取外界信息的主要來源，文本則是對事物的一種凝練描述，人通過眼睛捕獲文本獲取信息，機器設備的眼睛則是冰冷的攝像頭。如何讓機器設備從拍照獲取的圖像中準確檢測識別文本信息逐漸為各界學者關注。
現代文本檢測方法多為基于深度學習的方法，主要分為基于候選框和基于像素分割的兩種形式。本文選擇基于像素分割的深度學習模型作為文本檢測識別的主要研究方向，能夠同時滿足對自然場景文本的精確檢測，又能保證后續設備功能(如語義分析等功能)的拓展。
本文詳細內容請下載:http://www.cowatch.cn/resource/share/2000003385
作者信息:
陳小順，王良君
(江蘇大學 計算機科學與通信工程學院，江蘇 鎮江212013)
此內容為AET網站原創，未經授權禁止轉載。