Marmelator/layout-parser
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases 289 Wiki Activity

RED-9139: more robust TOC detection

Browse Source 
* detect numbers in words, and not just whole words that are numbers
This commit is contained in:
Kilian Schuettler 2024-11-08 10:41:56 +01:00
parent f9b25c8157
commit e6cd889444
9 changed files with 107 additions and 55 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
layoutparser-service/layoutparser-service-processor/src/main/java/com/knecon/fforesight/service/layoutparser/processor/model/SectionIdentifier.java
View File

@ -16,7 +16,7 @@ import lombok.experimental.FieldDefaults;
@FieldDefaults(makeFinal = true, level = AccessLevel.PRIVATE)
public class SectionIdentifier {
public static Pattern numericalIdentifierPattern = Pattern.compile("^[\\s]?(\\d{1,2})[\\s.,;]?(\\d{1,2})?[\\s.,;]?(\\d{1,2})?[\\s.,;]?(\\d{1,2})?");
public static Pattern numericalIdentifierPattern = Pattern.compile("^[\\s]?(\\d{1,2})(?:[\\s.,;](\\d{1,2}))?(?:[\\s.,;](\\d{1,2}))?(?:[\\s.,;](\\d{1,2}))?");
public static Pattern alphanumericIdentifierPattern = Pattern.compile("^[\\s]?[A-Za-z][\\s.,;]?(\\d{1,2})[\\s.,;]?(\\d{1,2})?[\\s.,;]?(\\d{1,2})?[\\s.,;]?(\\d{1,2})?[\\s.,;]?");
public enum Format {

16
layoutparser-service/layoutparser-service-processor/src/main/java/com/knecon/fforesight/service/layoutparser/processor/model/text/TextPositionSequenceComparator.java
View File

@ -3,19 +3,21 @@ package com.knecon.fforesight.service.layoutparser.processor.model.text;
import java.util.Comparator;
import java.util.HashMap;
public class TextPositionSequenceComparator implements Comparator<Word> {
import com.knecon.fforesight.service.layoutparser.processor.services.classification.NumberWord;
private HashMap<Word, TextBlockOnPage> lookup;
public class TextPositionSequenceComparator implements Comparator<NumberWord> {
private HashMap<NumberWord, TextBlockOnPage> lookup;
public TextPositionSequenceComparator(HashMap<Word, TextBlockOnPage> lookup) {
public TextPositionSequenceComparator(HashMap<NumberWord, TextBlockOnPage> lookup) {
this.lookup = lookup;
}
@Override
public int compare(Word number1, Word number2) {
public int compare(NumberWord number1, NumberWord number2) {
int page1 = lookup.get(number1).page().getPageNumber();
int page2 = lookup.get(number2).page().getPageNumber();
@ -24,11 +26,11 @@ public class TextPositionSequenceComparator implements Comparator<Word> {
return Integer.compare(page1, page2);
}
if (number1.getY() != number2.getY()) {
return Double.compare(number1.getY(), number2.getY());
if (number1.word().getY() != number2.word().getY()) {
return Double.compare(number1.word().getY(), number2.word().getY());
}
return Integer.compare(Integer.parseInt(number1.toString()), Integer.parseInt(number2.toString()));
return Integer.compare(number1.number(), number2.number());
}
}

7
layoutparser-service/layoutparser-service-processor/src/main/java/com/knecon/fforesight/service/layoutparser/processor/services/classification/NumberWord.java Normal file
View File

@ -0,0 +1,7 @@
package com.knecon.fforesight.service.layoutparser.processor.services.classification;
import com.knecon.fforesight.service.layoutparser.processor.model.text.Word;
public record NumberWord(Word word, int number) {
}

92
layoutparser-service/layoutparser-service-processor/src/main/java/com/knecon/fforesight/service/layoutparser/processor/services/classification/TableOfContentsClassificationService.java
View File

@ -14,6 +14,7 @@ import java.util.List;
import java.util.Locale;
import java.util.Map;
import java.util.Set;
import java.util.regex.Matcher;
import java.util.stream.Collectors;
import org.springframework.stereotype.Service;
@ -23,6 +24,7 @@ import com.knecon.fforesight.service.layoutparser.processor.model.AbstractPageBl
import com.knecon.fforesight.service.layoutparser.processor.model.ClassificationDocument;
import com.knecon.fforesight.service.layoutparser.processor.model.ClassificationPage;
import com.knecon.fforesight.service.layoutparser.processor.model.PageBlockType;
import com.knecon.fforesight.service.layoutparser.processor.model.SectionIdentifier;
import com.knecon.fforesight.service.layoutparser.processor.model.text.TextBlockOnPage;
import com.knecon.fforesight.service.layoutparser.processor.model.text.TextPageBlock;
import com.knecon.fforesight.service.layoutparser.processor.model.text.Word;
@ -71,9 +73,9 @@ public class TableOfContentsClassificationService {
ClassificationPage startPage = textBlocks.get(start).page();
List<TextBlockOnPage> initialLookAhead = textBlocks.subList(start, Math.min(start + SURROUNDING_BLOCKS_RADIUS, textBlocks.size()));
HashMap<Word, TextBlockOnPage> lookup = new HashMap<>();
List<Word> numbers = extractNumbers(initialLookAhead, lookup, document.getPages().size());
TocNumberFinder tocNumberFinder = new TocNumberFinder(numbers, lookup);
HashMap<NumberWord, TextBlockOnPage> numberToBlockLookup = new HashMap<>();
List<NumberWord> numbers = extractNumbers(initialLookAhead, numberToBlockLookup, document.getPages().size());
TocNumberFinder tocNumberFinder = new TocNumberFinder(numbers, numberToBlockLookup);
int lastCandidate = start;
for (int i = start; i < Math.min(lastCandidate + SURROUNDING_BLOCKS_RADIUS, textBlocks.size()); i++) {
@ -93,28 +95,28 @@ public class TableOfContentsClassificationService {
break;
}
List<Word> numbersFromBlock = extractNumbers(textBlockOnPage, lookup, document.getPages().size());
List<NumberWord> numbersFromBlock = extractNumbers(textBlockOnPage, numberToBlockLookup, document.getPages().size());
List<Word> currentRightmostCluster = tocNumberFinder.getCurrentRightmostCluster();
List<NumberWord> currentRightmostCluster = tocNumberFinder.getCurrentRightmostCluster();
if (currentRightmostCluster.size() < MINIMUM_MATCHES) {
log.debug("No numbers indicating a table of contents here.");
return start;
}
if (anyIntersection(currentRightmostCluster, numbersFromBlock, lookup)) {
if (anyIntersection(currentRightmostCluster, numbersFromBlock, numberToBlockLookup)) {
lastCandidate = i;
numbersFromBlock.forEach(tocNumberFinder::add);
}
}
addVisualization(document.getLayoutDebugLayer(), tocNumberFinder, lookup);
Set<TextBlockOnPage> blocksWithNumberInCluster = tocNumberFinder.getCurrentRightmostCluster()
.stream()
.map(lookup::get)
.map(numberToBlockLookup::get)
.collect(Collectors.toSet());
addVisualization(document.getLayoutDebugLayer(), tocNumberFinder, numberToBlockLookup, blocksWithNumberInCluster, textBlocks.get(start - 1));
int lastConfirmed = start;
for (int i = start; i < lastCandidate + 1; i++) {
TextBlockOnPage textBlockOnPage = textBlocks.get(i);
@ -132,18 +134,22 @@ public class TableOfContentsClassificationService {
}
private static void addVisualization(LayoutDebugLayer layoutDebugLayer, TocNumberFinder tocNumberFinder, Map<Word, TextBlockOnPage> lookup) {
private static void addVisualization(LayoutDebugLayer layoutDebugLayer,
TocNumberFinder tocNumberFinder,
Map<NumberWord, TextBlockOnPage> lookup,
Set<TextBlockOnPage> blocksWithNumberInCluster,
TextBlockOnPage startingHeadline) {
tocNumberFinder.getCurrentRightmostCluster()
.stream()
.collect(Collectors.groupingBy(key -> lookup.get(key).page().getPageNumber()))
.forEach((pageNumber, number) -> layoutDebugLayer.addTocPages(number, pageNumber));
layoutDebugLayer.addTocBlocks(blocksWithNumberInCluster);
layoutDebugLayer.addTocBlocks(Set.of(startingHeadline));
}
private static boolean anyIntersection(Collection<Word> numbers1,
Collection<Word> numbers2,
Map<Word, TextBlockOnPage> lookup) {
private static boolean anyIntersection(Collection<NumberWord> numbers1, Collection<NumberWord> numbers2, Map<NumberWord, TextBlockOnPage> lookup) {
return numbers1.stream()
.anyMatch(numberFromCluster -> numbers2.stream()
@ -151,9 +157,9 @@ public class TableOfContentsClassificationService {
}
private static List<Word> extractNumbers(List<TextBlockOnPage> textBlocks, Map<Word, TextBlockOnPage> lookup, int numberOfPages) {
private static List<NumberWord> extractNumbers(List<TextBlockOnPage> textBlocks, Map<NumberWord, TextBlockOnPage> lookup, int numberOfPages) {
List<Word> blocks = new LinkedList<>();
List<NumberWord> blocks = new LinkedList<>();
for (TextBlockOnPage textBlock : textBlocks) {
blocks.addAll(extractNumbers(textBlock, lookup, numberOfPages));
}
@ -161,30 +167,36 @@ public class TableOfContentsClassificationService {
}
private static List<Word> extractNumbers(TextBlockOnPage textBlock, Map<Word, TextBlockOnPage> lookup, int numberOfPages) {
private static List<NumberWord> extractNumbers(TextBlockOnPage textBlock, Map<NumberWord, TextBlockOnPage> lookup, int numberOfPages) {
List<Word> blocks = new LinkedList<>();
List<NumberWord> blocks = new LinkedList<>();
TextPageBlock block = textBlock.textBlock();
List<Word> sequences = block.getWords();
for (int i = 0; i < sequences.size(); i++) {
List<Word> words = block.getWords();
for (int i = 0; i < words.size(); i++) {
Word word = sequences.get(i);
Word word = words.get(i);
if (!NUMERIC.matcher(word).matches() || word.length() > 5) {
if (AMOUNT_PATTERN.matcher(getSurroundingString(i, words)).matches()) {
continue;
}
if (AMOUNT_PATTERN.matcher(getSurroundingString(i, sequences)).matches()) {
if (SectionIdentifier.fromSearchText(word.toString()).level() > 1) {
continue;
}
Matcher numberFinder = NUMERIC.matcher(word);
if (!numberFinder.find() || word.length() > 5) {
continue;
}
try {
int pageNumber = Integer.parseInt(word.toString());
int pageNumber = Integer.parseInt(numberFinder.group());
if (0 >= pageNumber || pageNumber > numberOfPages) {
continue;
}
lookup.put(word, textBlock);
blocks.add(word);
NumberWord numberWord = new NumberWord(word, pageNumber);
lookup.put(numberWord, textBlock);
blocks.add(numberWord);
} catch (NumberFormatException e) {
log.debug("That wasn't a number! Should not happen, due to numeric check beforehand.");
}
@ -203,13 +215,13 @@ public class TableOfContentsClassificationService {
}
private static boolean matches(Word number1, Word number2, Map<Word, TextBlockOnPage> lookup) {
private static boolean matches(NumberWord number1, NumberWord number2, Map<NumberWord, TextBlockOnPage> lookup) {
if (number1.getDir() != number2.getDir()) {
if (number1.word().getDir() != number2.word().getDir()) {
return false;
}
return number1.intersectsXDirAdj(number2, INTERSECTION_TOLERANCE);
return number1.word().intersectsXDirAdj(number2.word(), INTERSECTION_TOLERANCE);
}
@ -247,11 +259,11 @@ public class TableOfContentsClassificationService {
private static class TocNumberFinder {
final UnionFind<Word> numberClusters;
final HashMap<Word, TextBlockOnPage> lookup;
final UnionFind<NumberWord> numberClusters;
final HashMap<NumberWord, TextBlockOnPage> lookup;
TocNumberFinder(List<Word> blocks, HashMap<Word, TextBlockOnPage> lookup) {
TocNumberFinder(List<NumberWord> blocks, HashMap<NumberWord, TextBlockOnPage> lookup) {
this.numberClusters = new UnionFind<>(new HashSet<>(blocks));
for (int i = 0; i < blocks.size(); i++) {
@ -265,14 +277,14 @@ public class TableOfContentsClassificationService {
}
public void add(Word number) {
public void add(NumberWord number) {
if (numberClusters.getElements().contains(number)) {
return;
}
numberClusters.addElement(number);
for (Word element : numberClusters.getElements()) {
for (NumberWord element : numberClusters.getElements()) {
if (matches(number, element, lookup)) {
numberClusters.union(element, number);
}
@ -280,7 +292,7 @@ public class TableOfContentsClassificationService {
}
public List<Word> getCurrentRightmostCluster() {
public List<NumberWord> getCurrentRightmostCluster() {
return numberClusters.getGroups()
.stream()
@ -291,7 +303,7 @@ public class TableOfContentsClassificationService {
.map(this::removeOutliers)
// .map(this::filterByMinimumDensity)
.filter(cluster -> cluster.size() > MINIMUM_MATCHES)
.max(Comparator.comparingDouble(cluster -> cluster.get(0).getBBox().getMaxX())).orElse(Collections.emptyList());
.max(Comparator.comparingDouble(cluster -> cluster.get(0).word().getBBox().getMaxX())).orElse(Collections.emptyList());
}
// does not seem to be doing much, ideally instead of using the height of the blocks i would like to use the height, beginning from the MainBody top,
@ -322,9 +334,9 @@ public class TableOfContentsClassificationService {
// }
public List<Word> removeOutliers(List<Word> numbers) {
public List<NumberWord> removeOutliers(List<NumberWord> numbers) {
List<Word> result = new ArrayList<>();
List<NumberWord> result = new ArrayList<>();
result.add(numbers.get(0));
@ -346,7 +358,7 @@ public class TableOfContentsClassificationService {
// Helper method to check if removing the current number results in a better order
public static boolean isBetterWithout(List<Word> numbers, int i) {
public static boolean isBetterWithout(List<NumberWord> numbers, int i) {
if (i == 0 || i == numbers.size() - 1) {
return false;
@ -362,9 +374,9 @@ public class TableOfContentsClassificationService {
}
private static int getNumberAsInt(List<Word> numbers, int i) {
private static int getNumberAsInt(List<NumberWord> numbers, int i) {
return Integer.parseInt(numbers.get(i).toString());
return numbers.get(i).number();
}
}

8
layoutparser-service/layoutparser-service-processor/src/main/java/com/knecon/fforesight/service/layoutparser/processor/services/factory/DocumentGraphFactory.java
View File

@ -16,7 +16,7 @@ import java.util.Optional;
import java.util.Set;
import java.util.stream.Collectors;
import com.knecon.fforesight.service.layoutparser.internal.api.data.redaction.NodeType;
import com.knecon.fforesight.service.layoutparser.internal.api.data.redaction.NodeTypeProto;
import com.knecon.fforesight.service.layoutparser.internal.api.queue.LayoutParsingType;
import com.knecon.fforesight.service.layoutparser.processor.model.AbstractPageBlock;
import com.knecon.fforesight.service.layoutparser.processor.model.ClassificationDocument;
@ -82,9 +82,9 @@ public class DocumentGraphFactory {
documentGraph.streamAllSubNodes()
.filter(SemanticNode::isLeaf)
.filter(node -> !node.getType().equals(NodeType.HEADER))
.filter(node -> !node.getType().equals(NodeType.FOOTER))
.filter(node -> !node.getType().equals(NodeType.IMAGE))
.filter(node -> !node.getType().equals(NodeTypeProto.NodeType.HEADER))
.filter(node -> !node.getType().equals(NodeTypeProto.NodeType.FOOTER))
.filter(node -> !node.getType().equals(NodeTypeProto.NodeType.IMAGE))
.map(SemanticNode::getTextBlock)
.map(TextBlock::getAtomicTextBlocks)
.flatMap(Collection::stream)

25
layoutparser-service/layoutparser-service-processor/src/main/java/com/knecon/fforesight/service/layoutparser/processor/visualization/LayoutDebugLayer.java
View File

@ -27,7 +27,9 @@ import com.knecon.fforesight.service.layoutparser.processor.model.outline.Outlin
import com.knecon.fforesight.service.layoutparser.processor.model.table.CleanRulings;
import com.knecon.fforesight.service.layoutparser.processor.model.table.Ruling;
import com.knecon.fforesight.service.layoutparser.processor.model.text.ListIdentifier;
import com.knecon.fforesight.service.layoutparser.processor.model.text.TextBlockOnPage;
import com.knecon.fforesight.service.layoutparser.processor.model.text.Word;
import com.knecon.fforesight.service.layoutparser.processor.services.classification.NumberWord;
import com.knecon.fforesight.service.layoutparser.processor.utils.CoordinateTransforms;
import com.knecon.fforesight.service.layoutparser.processor.utils.MarkedContentUtils;
import com.knecon.fforesight.service.layoutparser.processor.utils.PageInformation;
@ -293,7 +295,7 @@ public class LayoutDebugLayer extends LayoutDebugLayerConfig {
}
public void addTocPages(List<Word> numbers, int page) {
public void addTocPages(List<NumberWord> numbers, int page) {
if (!active) {
return;
@ -302,11 +304,13 @@ public class LayoutDebugLayer extends LayoutDebugLayerConfig {
VisualizationsOnPage visualizationsOnPage = getOrCreateVisualizationsOnPage(page, this.tocPages);
visualizationsOnPage.getColoredRectangles()
.addAll(numbers.stream()
.map(NumberWord::word)
.map(BoundingBox::getBBoxPdf)
.map(line -> new ColoredRectangle(line, LINES_COLOR, LINE_WIDTH))
.toList());
visualizationsOnPage.getColoredRectangles()
.add(new ColoredRectangle(numbers.stream()
.map(NumberWord::word)
.map(BoundingBox::getBBoxPdf)
.collect(RectangleTransformations.collectBBox()), LINES_COLOR, LINE_WIDTH));
}
@ -332,8 +336,10 @@ public class LayoutDebugLayer extends LayoutDebugLayerConfig {
private void addOutlineObject(OutlineObject outlineObject, PageInformation pageInformation) {
if (!active) {
return;
}
int rectSize = 5;
Point2D point2D;
if (outlineObject.getPoint().isPresent()) {
point2D = outlineObject.getPoint().get();
@ -357,10 +363,25 @@ public class LayoutDebugLayer extends LayoutDebugLayerConfig {
public void addListIdentifiers(List<ListIdentifier> listIdentifiers) {
if (!active) {
return;
}
for (ListIdentifier listIdentifier : listIdentifiers) {
getOrCreateVisualizationsOnPage(listIdentifier.getPage(), this.listIdentifiers).getColoredRectangles()
.add(new ColoredRectangle(listIdentifier.getWord().getBBoxPdf(), WORDS_COLOR, LINE_WIDTH));
}
}
public void addTocBlocks(Set<TextBlockOnPage> blocksWithNumberInCluster) {
if (!active) {
return;
}
for (TextBlockOnPage textBlockOnPage : blocksWithNumberInCluster) {
getOrCreateVisualizationsOnPage(textBlockOnPage.page().getPageNumber(), this.tocBlocks).getColoredRectangles()
.add(new ColoredRectangle(textBlockOnPage.textBlock().getBBoxPdf(), TOC_COLOR, LINE_WIDTH));
}
}
}

7
layoutparser-service/layoutparser-service-server/src/test/java/com/knecon/fforesight/service/layoutparser/processor/model/SectionIdentifierTest.java
View File

@ -83,4 +83,11 @@ class SectionIdentifierTest {
assertEquals(List.of(4, 1, 2, 4), identifier.getIdentifiers());
}
@Test
void testFalsePositive111() {
SectionIdentifier identifier = SectionIdentifier.fromSearchText("111: Headline");
assertEquals(SectionIdentifier.Format.NUMERICAL, identifier.getFormat());
assertEquals(1, identifier.level());
}
}

1
layoutparser-service/viewer-doc-processor/src/main/java/com/knecon/fforesight/service/viewerdoc/LayerIdentifier.java
View File

@ -57,6 +57,7 @@ public record LayerIdentifier(String name, String markedContentName) {
public static final LayerIdentifier OUTLINE_OBJECTS = new LayerIdentifier("Outline Positions", "OUTLINE_OBJECTS");
public static final LayerIdentifier SENTENCES = new LayerIdentifier("Sentences", "SENTENCES");
public static final LayerIdentifier TOC_PAGES = new LayerIdentifier("TOC pages", "TOC_PAGES");
public static final LayerIdentifier TOC_BLOCKS = new LayerIdentifier("TOC blocks", "TOC_BLOCKS");
public static final LayerIdentifier LIST_IDENTIFIERS = new LayerIdentifier("List identifiers", "LIST_IDENTIFIERS");
// Visual layout parser

4
layoutparser-service/viewer-doc-processor/src/main/java/com/knecon/fforesight/service/viewerdoc/layers/LayoutDebugLayerConfig.java
View File

@ -22,6 +22,7 @@ public class LayoutDebugLayerConfig extends AbstractLayerGroup {
protected static final Color WORDS_COLOR = new Color(68, 84, 147);
protected static final Color LINES_COLOR = new Color(152, 45, 179);
protected static final Color TOC_COLOR = new Color(33, 159, 144);
protected static final Color ZONES_COLOR = new Color(131, 38, 38);
protected static final Color RULINGS_COLOR = new Color(21, 221, 174);
@ -59,6 +60,7 @@ public class LayoutDebugLayerConfig extends AbstractLayerGroup {
protected final Visualizations outlineObjects = Visualizations.builder().layer(LayerIdentifier.OUTLINE_OBJECTS).build();
protected final Visualizations sentences = Visualizations.builder().layer(LayerIdentifier.SENTENCES).build();
protected final Visualizations tocPages = Visualizations.builder().layer(LayerIdentifier.TOC_PAGES).build();
protected final Visualizations tocBlocks = Visualizations.builder().layer(LayerIdentifier.TOC_BLOCKS).build();
protected final Visualizations listIdentifiers = Visualizations.builder().layer(LayerIdentifier.LIST_IDENTIFIERS).build();
@ -77,9 +79,9 @@ public class LayoutDebugLayerConfig extends AbstractLayerGroup {
markedContent, //
outlineObjects, //
tocPages, //
tocBlocks, //
listIdentifiers //
);
}
}