如何使用pdfbox或其他Java库减小合并的PDF / A-1b文件的大小

Question

输入：包含嵌入式字体的（例如14个）PDF / A-1b文件列表。
处理：与Apache PDFBOX进行简单合并。
结果：1个PDF / A-1b文件，文件大小太大（太大）。（这几乎是所有源文件大小的总和）。

问题：是否可以减小生成的PDF的文件大小？
想法：删除多余的嵌入式字体。但是如何？这是正确的做法吗？

不幸的是，以下代码无法完成任务，但突出了明显的问题。

try (PDDocument document = PDDocument.load(new File("E:/tmp/16189_ZU_20181121195111_5544_2008-12-31_Standardauswertung.pdf"))) {
    List<COSName> collectedFonts = new ArrayList<>();
    PDPageTree pages = document.getDocumentCatalog().getPages();
    int pageNr = 0;
    for (PDPage page : pages) {
        pageNr++;
        Iterable<COSName> names = page.getResources().getFontNames();
        System.out.println("Page " + pageNr);
        for (COSName name : names) {
            collectedFonts.add(name);
            System.out.print("\t" + name + " - ");
            PDFont font = page.getResources().getFont(name);
            System.out.println(font + ", embedded: " + font.isEmbedded());
            page.getCOSObject().removeItem(COSName.F);
            page.getResources().getCOSObject().removeItem(name);
        }
    }
    document.save("E:/tmp/output.pdf");
}

该代码产生如下输出：

try (PDDocument document = PDDocument.load(new File("E:/tmp/16189_ZU_20181121195111_5544_2008-12-31_Standardauswertung.pdf"))) {
    List<COSName> collectedFonts = new ArrayList<>();
    PDPageTree pages = document.getDocumentCatalog().getPages();
    int pageNr = 0;
    for (PDPage page : pages) {
        pageNr++;
        Iterable<COSName> names = page.getResources().getFontNames();
        System.out.println("Page " + pageNr);
        for (COSName name : names) {
            collectedFonts.add(name);
            System.out.print("\t" + name + " - ");
            PDFont font = page.getResources().getFont(name);
            System.out.println(font + ", embedded: " + font.isEmbedded());
            page.getCOSObject().removeItem(COSName.F);
            page.getResources().getCOSObject().removeItem(name);
        }
    }
    document.save("E:/tmp/output.pdf");
}

任何帮助表示赞赏...

Answer 1

此答案中的代码是为了优化文档，例如OP的示例文档，即包含完全相同对象的副本的文档（在手头完全相同，完全嵌入的字体的情况下）的尝试。它不会仅将几乎相同的对象（例如，相同字体的多个子集）合并到一个单个联合子集中。

在对问题进行评论的过程中，很明显，OP的PDF中的重复字体确实是源字体文件的完全副本。为了合并这些重复的对象，必须收集文档的复杂对象（数组，字典，流），将它们相互比较，然后合并重复的对象。

由于在大型文档的情况下，文档的所有复杂对象的实际成对比较可能会花费太多时间，因此以下代码将计算这些对象的哈希值，并且仅比较具有相同哈希值的对象。

要合并重复项，代码将选择其中一个重复项，并将对任何其他重复项的所有引用替换为对所选重复项的引用，从而从文档对象池中删除其他重复项。为了更有效地执行此操作，代码最初不仅收集所有复杂对象，而且还收集每个对象的所有引用。

优化代码

这是用于优化的方法PDDocument：

public void optimize(PDDocument pdDocument) throws IOException {
    Map<COSBase, Collection<Reference>> complexObjects = findComplexObjects(pdDocument);
    for (int pass = 0; ; pass++) {
        int merges = mergeDuplicates(complexObjects);
        if (merges <= 0) {
            System.out.printf("Pass %d - No merged objects\n\n", pass);
            break;
        }
        System.out.printf("Pass %d - Merged objects: %d\n\n", pass, merges);
    }
}

（正在测试的OptimizeAfterMerge方法）

该优化需要多次通过，因为某些对象的相等性只能在它们引用的重复项合并后才能识别。

以下帮助器方法和类收集PDF的复杂对象以及对它们的引用：

Map<COSBase, Collection<Reference>> findComplexObjects(PDDocument pdDocument) {
    COSDictionary catalogDictionary = pdDocument.getDocumentCatalog().getCOSObject();
    Map<COSBase, Collection<Reference>> incomingReferences = new HashMap<>();
    incomingReferences.put(catalogDictionary, new ArrayList<>());

    Set<COSBase> lastPass = Collections.<COSBase>singleton(catalogDictionary);
    Set<COSBase> thisPass = new HashSet<>();
    while(!lastPass.isEmpty()) {
        for (COSBase object : lastPass) {
            if (object instanceof COSArray) {
                COSArray array = (COSArray) object;
                for (int i = 0; i < array.size(); i++) {
                    addTarget(new ArrayReference(array, i), incomingReferences, thisPass);
                }
            } else if (object instanceof COSDictionary) {
                COSDictionary dictionary = (COSDictionary) object;
                for (COSName key : dictionary.keySet()) {
                    addTarget(new DictionaryReference(dictionary, key), incomingReferences, thisPass);
                }
            }
        }
        lastPass = thisPass;
        thisPass = new HashSet<>();
    }
    return incomingReferences;
}

void addTarget(Reference reference, Map<COSBase, Collection<Reference>> incomingReferences, Set<COSBase> thisPass) {
    COSBase object = reference.getTo();
    if (object instanceof COSArray || object instanceof COSDictionary) {
        Collection<Reference> incoming = incomingReferences.get(object);
        if (incoming == null) {
            incoming = new ArrayList<>();
            incomingReferences.put(object, incoming);
            thisPass.add(object);
        }
        incoming.add(reference);
    }
}

（OptimizeAfterMerge帮助器方法findComplexObjects和addTarget）

interface Reference {
    public COSBase getFrom();

    public COSBase getTo();
    public void setTo(COSBase to);
}

static class ArrayReference implements Reference {
    public ArrayReference(COSArray array, int index) {
        this.from = array;
        this.index = index;
    }

    @Override
    public COSBase getFrom() {
        return from;
    }

    @Override
    public COSBase getTo() {
        return resolve(from.get(index));
    }

    @Override
    public void setTo(COSBase to) {
        from.set(index, to);
    }

    final COSArray from;
    final int index;
}

static class DictionaryReference implements Reference {
    public DictionaryReference(COSDictionary dictionary, COSName key) {
        this.from = dictionary;
        this.key = key;
    }

    @Override
    public COSBase getFrom() {
        return from;
    }

    @Override
    public COSBase getTo() {
        return resolve(from.getDictionaryObject(key));
    }

    @Override
    public void setTo(COSBase to) {
        from.setItem(key, to);
    }

    final COSDictionary from;
    final COSName key;
}

（OptimizeAfterMerge帮助器接口Reference与实现ArrayReference和DictionaryReference）

最后，以下辅助方法和类将识别并合并重复项：

int mergeDuplicates(Map<COSBase, Collection<Reference>> complexObjects) throws IOException {
    List<HashOfCOSBase> hashes = new ArrayList<>(complexObjects.size());
    for (COSBase object : complexObjects.keySet()) {
        hashes.add(new HashOfCOSBase(object));
    }
    Collections.sort(hashes);

    int removedDuplicates = 0;
    if (!hashes.isEmpty()) {
        int runStart = 0;
        int runHash = hashes.get(0).hash;
        for (int i = 1; i < hashes.size(); i++) {
            int hash = hashes.get(i).hash;
            if (hash != runHash) {
                int runSize = i - runStart;
                if (runSize != 1) {
                    System.out.printf("Equal hash %d for %d elements.\n", runHash, runSize);
                    removedDuplicates += mergeRun(complexObjects, hashes.subList(runStart, i));
                }
                runHash = hash;
                runStart = i;
            }
        }
        int runSize = hashes.size() - runStart;
        if (runSize != 1) {
            System.out.printf("Equal hash %d for %d elements.\n", runHash, runSize);
            removedDuplicates += mergeRun(complexObjects, hashes.subList(runStart, hashes.size()));
        }
    }
    return removedDuplicates;
}

int mergeRun(Map<COSBase, Collection<Reference>> complexObjects, List<HashOfCOSBase> run) {
    int removedDuplicates = 0;

    List<List<COSBase>> duplicateSets = new ArrayList<>();
    for (HashOfCOSBase entry : run) {
        COSBase element = entry.object;
        for (List<COSBase> duplicateSet : duplicateSets) {
            if (equals(element, duplicateSet.get(0))) {
                duplicateSet.add(element);
                element = null;
                break;
            }
        }
        if (element != null) {
            List<COSBase> duplicateSet = new ArrayList<>();
            duplicateSet.add(element);
            duplicateSets.add(duplicateSet);
        }
    }

    System.out.printf("Identified %d set(s) of identical objects in run.\n", duplicateSets.size());

    for (List<COSBase> duplicateSet : duplicateSets) {
        if (duplicateSet.size() > 1) {
            COSBase surviver = duplicateSet.remove(0);
            Collection<Reference> surviverReferences = complexObjects.get(surviver);
            for (COSBase object : duplicateSet) {
                Collection<Reference> references = complexObjects.get(object);
                for (Reference reference : references) {
                    reference.setTo(surviver);
                    surviverReferences.add(reference);
                }
                complexObjects.remove(object);
                removedDuplicates++;
            }
            surviver.setDirect(false);
        }
    }

    return removedDuplicates;
}

boolean equals(COSBase a, COSBase b) {
    if (a instanceof COSArray) {
        if (b instanceof COSArray) {
            COSArray aArray = (COSArray) a;
            COSArray bArray = (COSArray) b;
            if (aArray.size() == bArray.size()) {
                for (int i=0; i < aArray.size(); i++) {
                    if (!resolve(aArray.get(i)).equals(resolve(bArray.get(i))))
                        return false;
                }
                return true;
            }
        }
    } else if (a instanceof COSDictionary) {
        if (b instanceof COSDictionary) {
            COSDictionary aDict = (COSDictionary) a;
            COSDictionary bDict = (COSDictionary) b;
            Set<COSName> keys = aDict.keySet();
            if (keys.equals(bDict.keySet())) {
                for (COSName key : keys) {
                    if (!resolve(aDict.getItem(key)).equals(bDict.getItem(key)))
                        return false;
                }
                // In case of COSStreams we strictly speaking should
                // also compare the stream contents here. But apparently
                // their hashes coincide well enough for the original
                // hashing equality, so let's just assume...
                return true;
            }
        }
    }
    return false;
}

static COSBase resolve(COSBase object) {
    while (object instanceof COSObject)
        object = ((COSObject)object).getObject();
    return object;
}

（OptimizeAfterMerge辅助方法mergeDuplicates，mergeRun，equals，和resolve）

static class HashOfCOSBase implements Comparable<HashOfCOSBase> {
    public HashOfCOSBase(COSBase object) throws IOException {
        this.object = object;
        this.hash = calculateHash(object);
    }

    int calculateHash(COSBase object) throws IOException {
        if (object instanceof COSArray) {
            int result = 1;
            for (COSBase member : (COSArray)object)
                result = 31 * result + member.hashCode();
            return result;
        } else if (object instanceof COSDictionary) {
            int result = 3;
            for (Map.Entry<COSName, COSBase> entry : ((COSDictionary)object).entrySet())
                result += entry.hashCode();
            if (object instanceof COSStream) {
                try (   InputStream data = ((COSStream)object).createRawInputStream()   ) {
                    MessageDigest md = MessageDigest.getInstance("MD5");
                    byte[] buffer = new byte[8192];
                    int bytesRead = 0;
                    while((bytesRead = data.read(buffer)) >= 0)
                        md.update(buffer, 0, bytesRead);
                    result = 31 * result + Arrays.hashCode(md.digest());
                } catch (NoSuchAlgorithmException e) {
                    throw new IOException(e);
                }
            }
            return result;
        } else {
            throw new IllegalArgumentException(String.format("Unknown complex COSBase type %s", object.getClass().getName()));
        }
    }

    final COSBase object;
    final int hash;

    @Override
    public int compareTo(HashOfCOSBase o) {
        int result = Integer.compare(hash,  o.hash);
        if (result == 0)
            result = Integer.compare(hashCode(), o.hashCode());
        return result;
    }
}

（OptimizeAfterMerge帮助器类HashOfCOSBase）

将代码应用于OP的示例文档

OP的示例文档大小约为6.5 MB。像这样应用上面的代码

PDDocument pdDocument = PDDocument.load(SOURCE);

optimize(pdDocument);

pdDocument.save(RESULT);

导致PDF的大小小于700 KB，并且看起来是完整的。

（如果有什么遗漏，请告诉我，我会尽力解决。）

警告语

一方面，此优化器无法识别所有相同的重复项。特别是在循环引用的情况下，将不会识别对象的重复圆，因为该代码仅在其内容相同时才识别重复的对象，而重复内容通常不会出现在重复的对象圆中。

另一方面，在某些情况下，此优化器可能过于急切，因为可能需要将某些重复项作为单独的对象，以供PDF查看器将每个实例作为一个单独的实体接受。

此外，该计划倒是各类对象的文件中，即使是那些定义PDF的内部结构，但它不会尝试更新任何PDFBox的类管理这个结构（PDDocument，PDDocumentCatalog，PDAcroForm，...）。为了避免任何待处理的更改使整个文档搞乱，因此，请仅将此程序应用于新加载的未修改PDDocument实例，并在不久之后进行保存。

Answer 2

在文件中调试时，我发现相同字体的字体文件被多次引用。因此，将字典中的实际字体文件项替换为已查看的字体文件项，引用已被删除，并且可以进行压缩。通过那个，我能够将一个 30 MB 的文件缩小到大约 6 MB。

    File file = new File("test.pdf");

    PDDocument doc = PDDocument.load(file);
    Map<String, COSBase> fontFileCache = new HashMap<>();
    for (int pageNumber = 0; pageNumber < doc.getNumberOfPages(); pageNumber++) {
        final PDPage page = doc.getPage(pageNumber);
        COSDictionary pageDictionary = (COSDictionary) page.getResources().getCOSObject().getDictionaryObject(COSName.FONT);
        for (COSName currentFont : pageDictionary.keySet()) {
            COSDictionary fontDictionary = (COSDictionary) pageDictionary.getDictionaryObject(currentFont);
            for (COSName actualFont : fontDictionary.keySet()) {
                COSBase actualFontDictionaryObject = fontDictionary.getDictionaryObject(actualFont);
                if (actualFontDictionaryObject instanceof COSDictionary) {
                    COSDictionary fontFile = (COSDictionary) actualFontDictionaryObject;
                    if (fontFile.getItem(COSName.FONT_NAME) instanceof COSName) {
                        COSName fontName = (COSName) fontFile.getItem(COSName.FONT_NAME);
                        fontFileCache.computeIfAbsent(fontName.getName(), key -> fontFile.getItem(COSName.FONT_FILE2));
                        fontFile.setItem(COSName.FONT_FILE2, fontFileCache.get(fontName.getName()));
                    }
                }
            }
        }
    }

    final ByteArrayOutputStream baos = new ByteArrayOutputStream();
    doc.save(baos);
    final File compressed = new File("test_compressed.pdf");
    baos.writeTo(new FileOutputStream(compressed));

也许这不是最优雅的方法，但它可以工作并保持 PDF/A-1b 兼容性。