English | 简体中文
智能文本自动处理工具（Intelligent text automatic processing tool）。
AutoText的功能主要有文本纠错，图片ocr以及表格结构识别等。
Guide
- 文本纠错部分详细见jcorrector
- 本项目目前主要包括：
-
基于ngram的纠错
- 1.纠错
Corrector corrector = new Corrector(); String sentence = “少先队员因该为老人让坐”; System.out.println(corrector.correct(sentence));- 2.检测
Detector detector = new Detector(); String sentence = “我的喉咙发炎了要买点阿莫细林吉”; System.out.println(detector.detect(sentence)); -
基于深度学习的纠错
- 纠错
LoadModel.loadOnnxModel(); String text = "今天新情很好。"; text = "你找到你最喜欢的工作，我也很高心。"; text = "是的，当线程不再使用时，该缓冲区将被清理（我昨天实际上对此进行了测试，我可以每5ms发送一个新线程，而不会产生净内存累积，并确认它的rng内存已在gc上清理）。编号7788"; text = text.toLowerCase(); BertTokenizer tokenizer = new BertTokenizer(); MacBert macBert = new MacBert(tokenizer); Map<String, OnnxTensor> inputTensor = null; try { inputTensor = macBert.parseInputText(text); } catch (Exception e) { e.printStackTrace(); } List<String> predTokenList = macBert.predCSC(inputTensor); predTokenList = predTokenList.stream().map(token -> token.replace("##", "")).collect(Collectors.toList()); String predString = String.join("", predTokenList); System.out.println(predString); List<Pair<String, String>> resultList = macBert.getErrors(predString, text); for(Pair<String, String> result : resultList) { System.out.println(text + " => " + result.getLeft() + " " + result.getRight()); } -
基于模板中文语法纠错
- 纠错
String templatePath = GecDemo.class.getClassLoader().getResource(PropertiesReader.get("template")).getPath().replaceFirst("/", ""); GecCheck gecRun = new GecCheck(); gecRun.init(templatePath); String sentence = "爸爸看完小品后忍俊不禁笑了起来。"; String infoStr = gecRun.checkCorrect(sentence); if(StringUtils.isNotBlank(infoStr)) { System.out.println(infoStr); } -
成语、专名纠错
- 纠错
String properNamePath = ProperDemo.class.getClassLoader().getResource(PropertiesReader.get("proper_name_path")).getPath().replaceFirst("/", ""); String strokePath = ProperDemo.class.getClassLoader().getResource(PropertiesReader.get("stroke_path")).getPath().replaceFirst("/", "");; ProperCorrector properCorrector = new ProperCorrector(properNamePath, strokePath); List<String> testLine = List.of( "报应接中迩来", "这块名表带带相传", "这块名表代代相传", "他贰话不说把牛奶喝完了", "这场比赛我甘败下风", "这场比赛我甘拜下封", "这家伙还蛮格尽职守的", "报应接中迩来", // 接踵而来 "人群穿流不息", "这个消息不径而走", "这个消息不胫儿走", "眼前的场景美仑美幻简直超出了人类的想象", "看着这两个人谈笑风声我心理不由有些忌妒", "有了这一番旁证博引", "有了这一番旁针博引", "这群鸟儿迁洗到远方去了", "这群鸟儿千禧到远方去了", "美国前总统特琅普给普京点了一个赞，特朗普称普金做了一个果断的决定" ); for(String line : testLine) { System.out.println(properCorrector.correct(line)); }
-
- 具体使用见本项目中的examples/correct部分，或者jcorrector
-
这部分主要利用paddleocr 中的检测与识别部分，并将其中模型转为onnx格式进行调用，本项目在识别前对图片进行了预处理，使得在cpu环境下，平均一张图10秒左右。
-
具体使用见本项目中的examples/ocr/text/OcrDemo部分
-
PS
- 模型网盘下载
- 提取码：b5vq
- 模型下载后可放入resources的text_recgo下或其它位置
-
使用
// read image file
String imagePath = "examples\\ocr\\img_test\\text_example.png";
var imageFile = Paths.get(imagePath);
var image = ImageFactory.getInstance().fromFile(imageFile);
// init model
String detectionModelFile = OcrDemo.class.getClassLoader().getResource(PropertiesReader.get("text_recog_det_model_path")).getPath().replaceFirst("/", "");
String recognitionModelFile = OcrDemo.class.getClassLoader().getResource(PropertiesReader.get("text_recog_rec_model_path")).getPath().replaceFirst("/", "");
Path detectionModelPath = Paths.get(detectionModelFile);
Path recognitionModelPath = Paths.get(recognitionModelFile);
OcrApp ocrApp = new OcrApp(detectionModelPath, recognitionModelPath);
ocrApp.init();
// predict result and consume time
var timeInferStart = System.currentTimeMillis();
Pair<List<TextListBox>, Image> imagePair = ocrApp.ocrImage(image, 960);
System.out.println("consume time: " + (System.currentTimeMillis() - timeInferStart)/1000.0 + "s");
for (var result : imagePair.getLeft()) {
System.out.println(result);
}
// save ocr result image
ocrApp.saveImageOcrResult(imagePair, "ocr_result.png", "examples\\ocr\\output");
ocrApp.closeAllModel();- 结果，为文字及其坐标
position: [800.0, 609.0, 877.0, 609.0, 877.0, 645.0, 800.0, 645.0], text: 8.23%
position: [433.0, 607.0, 494.0, 607.0, 494.0, 649.0, 433.0, 649.0], text: 68.4
position: [96.0, 610.0, 316.0, 611.0, 316.0, 641.0, 96.0, 640.0], text: 股东权益比率（%）
position: [624.0, 605.0, 688.0, 605.0, 688.0, 650.0, 624.0, 650.0], text: 63.2
position: [791.0, 570.0, 887.0, 570.0, 887.0, 600.0, 791.0, 600.0], text: -39.64%
position: [625.0, 564.0, 687.0, 564.0, 687.0, 606.0, 625.0, 606.0], text: 49.7
position: [134.0, 568.0, 279.0, 568.0, 279.0, 598.0, 134.0, 598.0], text: 毛利率（%）
......
- 结果图片展示图片OCR
- 利用yolov8进行版面检测，见layout_analysis4j
- 具体使用见本项目中的examples/ocr/layout_detection/LayoutDetection部分
- 使用
public static void main(String...args) {
String modelPath = LayoutDetection.class.getClassLoader().getResource(PropertiesReader.get("model_path")).getPath().replaceFirst("/", "");
String labelPath = LayoutDetection.class.getClassLoader().getResource(PropertiesReader.get("table_det_labels_path")).getPath().replaceFirst("/", "");
String imgPath = "D:\\project\\idea_workspace\\layout_analysis4j\\img\\test.webp";
try {
LayoutDet modelDet = new LayoutDet(modelPath, labelPath);
Mat img = Imgcodecs.imread(imgPath);
if (img.dataAddr() == 0) {
System.out.println("Could not open image: " + imgPath);
System.exit(1);
}
// run detection
try {
List<Detection> detectionList = modelDet.detectObjects(img);
LayoutDetectionUtil.drawPredictions(img, detectionList);
System.out.println(JSON.toJSONString(detectionList));
Imgcodecs.imwrite("D:\\project\\idea_workspace\\layout_analysis4j\\img\\prediction.jpg", img);
} catch (OrtException ortException) {
ortException.printStackTrace();
}
} catch (OrtException e) {
e.printStackTrace();
}
}- 结果如下
[{"bbox":[137.88228,40.05045,352.5302,60.206684],"confidence":0.9228547,"label":"Header","labelIndex":0},
{"bbox":[25.661982,52.15992,80.54977,60.164627],"confidence":0.8484751,"label":"Header","labelIndex":0},
{"bbox":[400.68176,50.069782,462.38123,58.523815],"confidence":0.83252084,"label":"Header","labelIndex":0},
{"bbox":[27.056168,478.51273,205.72672,656.39886],"confidence":0.9614719,"label":"Text","labelIndex":1},
{"bbox":[25.820251,304.84778,463.41486,386.2965],"confidence":0.89359975,"label":"Text","labelIndex":1},
{"bbox":[21.327255,190.66518,463.8985,257.07446],"confidence":0.8879021,"label":"Text","labelIndex":1},
{"bbox":[182.88458,142.3864,308.64737,156.58653],"confidence":0.79081506,"label":"Text","labelIndex":1},
{"bbox":[38.471603,435.21515,463.1955,474.5235],"confidence":0.77674204,"label":"Text","labelIndex":1},
{"bbox":[153.92957,160.85332,338.4781,168.90303],"confidence":0.764402,"label":"Text","labelIndex":1},
{"bbox":[27.318249,661.32355,151.53812,670.04987],"confidence":0.3412643,"label":"Text","labelIndex":1},
{"bbox":[306.27896,667.1539,362.94162,674.0262],"confidence":0.8710417,"label":"Figure caption","labelIndex":3},
{"bbox":[213.9415,479.61642,468.25687,661.9558],"confidence":0.9372132,"label":"Figure","labelIndex":4},
{"bbox":[26.771957,405.50818,94.59786,416.935],"confidence":0.91822684,"label":"Title","labelIndex":7},
{"bbox":[131.77039,103.47922,359.43063,120.83272],"confidence":0.88686645,"label":"Title","labelIndex":7},
{"bbox":[26.655102,661.2926,151.92046,670.0917],"confidence":0.87808716,"label":"Title","labelIndex":7},
{"bbox":[27.927279,275.91486,68.040955,287.49615],"confidence":0.8072859,"label":"Title","labelIndex":7},
{"bbox":[27.4192,661.2635,151.4754,670.21075],"confidence":0.49235547,"label":"Footer","labelIndex":8}]
- 结果图片展示版面检测
- 基于规则由opencv研发，主要识别的表格类型有：有边界表格、无边界表格以及部分有边界表格。
- 具体使用见本项目中的examples/ocr/table/TableDemo部分
- 使用
public static void borderedRecog() {
String imagePath = "D:\\project\\idea_workspace\\AutoText\\src\\main\\java\\examples\\ocr\\img_test\\bordered_example.png";
Mat imageMat = imread(imagePath);
System.out.println("imageMat : " + imageMat.size().height() + " " + imageMat.size().width() + " ");
List<List<List<Integer>>> resultList = BorderedRecog.recognizeStructure(imageMat);
System.out.println(resultList);
// ImageUtils.imshow("Image", pair.getRight());
}
public static void unBorderedRecog() {
String imagePath = "D:\\project\\idea_workspace\\AutoText\\src\\main\\java\\examples\\ocr\\img_test\\unbordered_example.jpg";
Mat imageMat = imread(imagePath);
System.out.println("imageMat : " + imageMat.size().height() + " " + imageMat.size().width() + " ");
List<List<List<Integer>>> resultList = UnBorderedRecog.recognizeStructure(imageMat);
System.out.println(resultList);
// ImageUtils.imshow("Image", pair.getRight());
}
public static void partiallyBorderedRecog() {
String imagePath = "D:\\project\\idea_workspace\\AutoText\\src\\main\\java\\examples\\ocr\\img_test\\partially_example.jpg";
Mat imageMat = imread(imagePath);
System.out.println("imageMat : " + imageMat.size().height() + " " + imageMat.size().width() + " ");
List<List<List<Integer>>> resultList = PartiallyBorderedRecog.recognizeStructure(imageMat);
System.out.println(resultList);
// ImageUtils.imshow("Image", pair.getRight());
}- 结果，为表格单元格坐标
[[[58, 48, 247, 182], [560, 48, 247, 182], [811, 48, 246, 182], [309, 48, 247, 182], [1312, 48, 247, 182],
[1061, 48, 247, 182]], [[58, 234, 247, 118], [309, 234, 247, 118], [1061, 234, 247, 118], [560, 234, 247, 118],
[811, 234, 246, 118], [1312, 234, 247, 118]], [[58, 356, 247, 118], [309, 356, 247, 118], [560, 356, 247, 118],
[811, 356, 246, 118], [1061, 356, 247, 118], [1312, 356, 247, 118]], [[58, 478, 247, 118], [309, 478, 247, 118],
[560, 478, 247, 118], [811, 478, 246, 118], [1061, 478, 247, 118], [1312, 478, 247, 118]], [[58, 600, 247, 119],
[309, 600, 247, 119], [560, 600, 247, 119], [811, 600, 246, 119], [1061, 600, 247, 119], [1312, 600, 247, 119]],
[[58, 723, 247, 118], [309, 723, 247, 118], [560, 723, 247, 118], [1061, 723, 247, 118], [1312, 723, 247, 118],
[811, 723, 246, 118]], [[58, 845, 247, 118], [309, 845, 247, 118], [560, 845, 247, 118], [811, 845, 246, 118],
[1312, 845, 247, 118], [1061, 845, 247, 118]]]
- 结果图片展示表格结构
- 这部分将整合表格结构和OCR识别，同时识别出表格单元格和OCR文本。
- 具体使用见本项目中的examples/ocr/table_text/TableTextDemo部分
- 使用
public static void main(String...args) throws IOException, TranslateException {
String imagePath = "D:\\project\\idea_workspace\\AutoText\\src\\main\\java\\examples\\ocr\\img_test\\bordered_example.png";
TableText tableText = new TableText();
/**
* maxSideLen:image resize
*
* borderType:{0:all, 1:bordered(default), 2:unbordered, 3:partiallybordered}
*/
int maxSideLen = -1; // default, no resize
int borderType = 1; // default, bordered
List<TextListBox> listBoxes = tableText.tableTextRecog(imagePath);
for(TextListBox textListBox : listBoxes) {
System.out.print(textListBox);
}
}- 结果，为表格单元格坐标以及单元格内的文本
position: [58.0, 48.0, 305.0, 48.0, 305.0, 230.0, 58.0, 230.0], text: 节次 星期
position: [309.0, 48.0, 556.0, 48.0, 556.0, 230.0, 309.0, 230.0], text: 周一
position: [811.0, 48.0, 1057.0, 48.0, 1057.0, 230.0, 811.0, 230.0], text: 周三
position: [1061.0, 48.0, 1308.0, 48.0, 1308.0, 230.0, 1061.0, 230.0], text: 周四
position: [560.0, 48.0, 807.0, 48.0, 807.0, 230.0, 560.0, 230.0], text: 周二
position: [1312.0, 48.0, 1559.0, 48.0, 1559.0, 230.0, 1312.0, 230.0], text: 周五
position: [58.0, 234.0, 305.0, 234.0, 305.0, 352.0, 58.0, 352.0], text:
position: [309.0, 234.0, 556.0, 234.0, 556.0, 352.0, 309.0, 352.0], text: 语文
position: [811.0, 234.0, 1057.0, 234.0, 1057.0, 352.0, 811.0, 352.0], text: 英语
position: [560.0, 234.0, 807.0, 234.0, 807.0, 352.0, 560.0, 352.0], text: 英语
position: [1061.0, 234.0, 1308.0, 234.0, 1308.0, 352.0, 1061.0, 352.0], text: 自然
position: [1312.0, 234.0, 1559.0, 234.0, 1559.0, 352.0, 1312.0, 352.0], text: 数学
position: [58.0, 356.0, 305.0, 356.0, 305.0, 474.0, 58.0, 474.0], text: 3
position: [560.0, 356.0, 807.0, 356.0, 807.0, 474.0, 560.0, 474.0], text: 英语
position: [309.0, 356.0, 556.0, 356.0, 556.0, 474.0, 309.0, 474.0], text: 语文
position: [811.0, 356.0, 1057.0, 356.0, 1057.0, 474.0, 811.0, 474.0], text: 英语
position: [1312.0, 356.0, 1559.0, 356.0, 1559.0, 474.0, 1312.0, 474.0], text: 数学
position: [1061.0, 356.0, 1308.0, 356.0, 1308.0, 474.0, 1061.0, 474.0], text: 语文
position: [58.0, 478.0, 305.0, 478.0, 305.0, 596.0, 58.0, 596.0], text: 三
position: [309.0, 478.0, 556.0, 478.0, 556.0, 596.0, 309.0, 596.0], text: 数学
position: [560.0, 478.0, 807.0, 478.0, 807.0, 596.0, 560.0, 596.0], text: 语文
position: [811.0, 478.0, 1057.0, 478.0, 1057.0, 596.0, 811.0, 596.0], text: 数学
position: [1312.0, 478.0, 1559.0, 478.0, 1559.0, 596.0, 1312.0, 596.0], text: 英语
position: [1061.0, 478.0, 1308.0, 478.0, 1308.0, 596.0, 1061.0, 596.0], text: 语文
position: [58.0, 600.0, 305.0, 600.0, 305.0, 719.0, 58.0, 719.0], text: 四
position: [309.0, 600.0, 556.0, 600.0, 556.0, 719.0, 309.0, 719.0], text: 数学
position: [811.0, 600.0, 1057.0, 600.0, 1057.0, 719.0, 811.0, 719.0], text: 数学
position: [560.0, 600.0, 807.0, 600.0, 807.0, 719.0, 560.0, 719.0], text: 语文
position: [1061.0, 600.0, 1308.0, 600.0, 1308.0, 719.0, 1061.0, 719.0], text: 体育
position: [1312.0, 600.0, 1559.0, 600.0, 1559.0, 719.0, 1312.0, 719.0], text: 英语
position: [58.0, 723.0, 305.0, 723.0, 305.0, 841.0, 58.0, 841.0], text: 五
position: [560.0, 723.0, 807.0, 723.0, 807.0, 841.0, 560.0, 841.0], text: 思想品德
position: [309.0, 723.0, 556.0, 723.0, 556.0, 841.0, 309.0, 841.0], text: 体育
position: [1061.0, 723.0, 1308.0, 723.0, 1308.0, 841.0, 1061.0, 841.0], text: 数学
position: [1312.0, 723.0, 1559.0, 723.0, 1559.0, 841.0, 1312.0, 841.0], text: 手工
position: [811.0, 723.0, 1057.0, 723.0, 1057.0, 841.0, 811.0, 841.0], text: 语文
position: [58.0, 845.0, 305.0, 845.0, 305.0, 963.0, 58.0, 963.0], text: 六
position: [309.0, 845.0, 556.0, 845.0, 556.0, 963.0, 309.0, 963.0], text: 美术
position: [560.0, 845.0, 807.0, 845.0, 807.0, 963.0, 560.0, 963.0], text: 音乐
position: [1061.0, 845.0, 1308.0, 845.0, 1308.0, 963.0, 1061.0, 963.0], text: 数学
position: [811.0, 845.0, 1057.0, 845.0, 1057.0, 963.0, 811.0, 963.0], text: 语文
position: [1312.0, 845.0, 1559.0, 845.0, 1559.0, 963.0, 1312.0, 963.0], text: 写字
- 加入jcorrector文本纠错，修改部分程序
- 基于paddleocr模型，利用java实现图片ocr
- 基于规则利用opencv识别表格结构
- 整合规则表格识别与OCR识别
- 加入版面检测分析
...
1、github：https://github.com/jiangnanboy
2、博客：https://www.cnblogs.com/little-horse/
3、邮件:2229029156@qq.com
如果你在研究中使用了AutoText，请按如下格式引用：
@{AutoText,
author = {Shi Yan},
title = {AutoText: Text automatic processing tool},
year = {2023},
url = {https://github.com/jiangnanboy/AutoText},
}AutoText 的授权协议为 Apache License 2.0，可免费用做商业用途。请在产品说明中附加AutoText的链接和授权协议。AutoText受版权法保护，侵权必究。
欢迎有兴趣的朋友fork，star，提交PR。