SPB Git

Repository Public License

OCRAndDocxTests.swift 12.8 KB 180 lines swift

1// ORIX Convert — Packages/EnginePDF/Tests/EnginePDFTests/OCRAndDocxTests.swift2// Description: Searchable PDF (invisible text layer positions, structure preservation) and PDF → DOCX / RTF.3// Author: Simon-Pierre Boucher — contact@spboucher.ai4import Foundation5import AppKit6import CoreGraphics7import PDFKit8import Testing9import ConvertCore10@testable import EnginePDF1112/// Runs a body that needs Vision; an unavailable recogniser is reported as a known issue instead of a failure.13func withVision(_ body: () async throws -> Void) async throws {14    do { try await body() } catch ConversionError.engineUnavailable(let reason) {15        withKnownIssue("Vision text recognition unavailable: \(reason)") { Issue.record("skipped") }16    }17}1819extension PDFEngineTests {20    @Suite struct SearchablePDF {21        /// Word boxes of the invisible layer against the boxes the fixture drew, on an upright page, a /Rotate 9022        /// page and a page whose media box is offset with a smaller crop box.23        @Test(arguments: [150.0, 300.0])24        func invisibleLayerMatchesTheScan(dpi: Double) async throws {25            try await withVision {26                let pdf = Fx.scannedPDF("scan-\(Int(dpi)).pdf", dpi: dpi, variants: true)27                let result = try await Fx.run(Fx.request(pdf, to: .pdf, op: PDFEngine.Operation.ocr))28                #expect(result.output.facts["ocr.pages"] == "3")29                let source = try #require(PDFDocument(url: pdf)), output = try #require(PDFDocument(url: result.file))30                #expect(output.pageCount == 3)31                let expected = Fx.scanProbeRects()32                var worst: CGFloat = 0, worstX: CGFloat = 0, worstY: CGFloat = 033                for index in 0..<3 {34                    let page = try #require(output.page(at: index)), original = try #require(source.page(at: index))35                    #expect(page.rotation == original.rotation)36                    #expect(page.bounds(for: .mediaBox) == original.bounds(for: .mediaBox))37                    #expect(page.bounds(for: .cropBox) == original.bounds(for: .cropBox))38                    let text = page.string ?? ""39                    for probe in Fx.scanProbes {40                        #expect(text.contains(probe), "page \(index + 1) lacks \(probe) at \(Int(dpi)) dpi: \(text)")41                        let selection = try #require(output.findString(probe, withOptions: []).first { $0.pages.first == page }, "no selection for \(probe) on page \(index + 1)")42                        let found = selection.bounds(for: page), want = Fx.expectedRect(expected[probe]!, page: index)43                        let overlap = found.intersection(want)44                        let covered = overlap.isNull ? 0 : (overlap.width * overlap.height) / (want.width * want.height)45                        #expect(covered > 0.6, "\(probe) page \(index + 1): found \(found), expected \(want)")46                        let drift = max(abs(found.midX - want.midX), abs(found.midY - want.midY))47                        worst = max(worst, drift); worstX = max(worstX, abs(found.midX - want.midX)); worstY = max(worstY, abs(found.midY - want.midY))48                        #expect(drift < 7, "\(probe) page \(index + 1): centre off by \(drift) pt")49                    }50                    // The scan itself is carried over, not re-encoded.51                    let before = ImageCollector(), after = ImageCollector()52                    before.collect(page: original.pageRef!, number: 1); after.collect(page: page.pageRef!, number: 1)53                    #expect(after.images.count == 1)54                    #expect(CGPDFRead.int(after.images[0].dictionary, "Width") == CGPDFRead.int(before.images[0].dictionary, "Width"))55                    #expect(CGPDFRead.filters(after.images[0].dictionary) == CGPDFRead.filters(before.images[0].dictionary))56                }57                // A media box that does not start at the origin, with a smaller crop box.58                let offset = Fx.offsetScanPDF("offset-\(Int(dpi)).pdf", dpi: dpi)59                let shifted = try await Fx.run(Fx.request(offset, to: .pdf, op: PDFEngine.Operation.ocr))60                let shiftedDocument = try #require(PDFDocument(url: shifted.file)), shiftedPage = try #require(shiftedDocument.page(at: 0))61                // PDFKit rewrites the boxes relative to the media-box corner: same sizes, origin moved to (0, 0).62                #expect(shiftedPage.bounds(for: .mediaBox) == CGRect(x: 0, y: 0, width: 612, height: 792))63                #expect(shiftedPage.bounds(for: .cropBox) == CGRect(x: 10, y: 10, width: 592, height: 772))64                for probe in Fx.scanProbes {65                    let selection = try #require(shiftedDocument.findString(probe, withOptions: []).first, "no selection for \(probe) on the offset page")66                    let found = selection.bounds(for: shiftedPage), want = expected[probe]!67                    let drift = max(abs(found.midX - want.midX), abs(found.midY - want.midY))68                    worst = max(worst, drift)69                    #expect(drift < 7, "\(probe) offset page: found \(found), expected \(want)")70                }71                print("MEASURE ocr-layer dpi=\(Int(dpi)) worst-centre-drift=\(String(format: "%.2f", worst))pt along-page-x=\(String(format: "%.2f", worstX))pt along-page-y=\(String(format: "%.2f", worstY))pt")72                // Words are separate in the extracted text.73                #expect((output.page(at: 0)?.string ?? "").contains("Quantum physics describes nature"))74            }75        }7677        @Test func outlineLinksAnnotationsAndFormsSurvive() async throws {78            try await withVision {79                let merged = try await Fx.run(Fx.request(Fx.richPDF("ocr-rich.pdf"), to: .pdf, op: PDFEngine.Operation.merge, more: [Fx.scannedPDF("ocr-scan.pdf", dpi: 200)]))80                let before = Fx.structure(merged.file)81                #expect(before.pages == 5 && before.outline == 6 && before.links == 2 && before.widgets == 1 && before.squares == 1)82                let result = try await Fx.run(Fx.request(merged.file, to: .pdf, op: PDFEngine.Operation.ocr))83                let after = Fx.structure(result.file)84                #expect(after == before)85                #expect(result.output.facts["ocr.pages"] == "1")86                #expect(Fx.pdfText(result.file, page: 4).contains("Quantum"))87                #expect(Fx.pdfText(result.file, page: 0) == Fx.pdfText(merged.file, page: 0))88                let info = PDFInspector.infoDictionary(result.file, password: nil)89                #expect(info["Title"] == "Rich Fixture" && info["Author"] == "ORIX Tests")90                // Annotations stay annotations: they are not painted into the page content.91                let flat = try await Fx.run(Fx.request(result.file, to: .png, ["pages": .string("3"), "resolution.dpi": .int(72), "annotations": .bool(false)]))92                let pixel = Fx.pixel(try #require(Fx.load(flat.file)), 400, 442)93                #expect(pixel.g > 240)94                let reopened = try #require(PDFDocument(url: result.file)), outline = try #require(reopened.outlineRoot)95                #expect(outline.child(at: 0)?.label == "ocr-rich" && outline.child(at: 0)?.child(at: 1)?.child(at: 0)?.label == "Section 2.1")96            }97        }9899        @Test func documentsWithTextAreReturnedUnchanged() async throws {100            let pdf = Fx.simplePDF("has-text.pdf", pages: 2)101            let result = try await Fx.run(Fx.request(pdf, to: .pdf, op: PDFEngine.Operation.ocr))102            #expect(result.output.facts["result"] == "unchanged" && result.hasWarning("ocr-nothing-added"))103            #expect(try Data(contentsOf: result.file) == Data(contentsOf: pdf))104            #expect(Fx.engine.preflight(Fx.request(pdf, to: .pdf, op: PDFEngine.Operation.ocr), inspection: nil).contains { $0.code == "ocr-not-needed" })105        }106107        @Test func everyPageOnRequest() async throws {108            try await withVision {109                let pdf = Fx.simplePDF("ocr-all.pdf", pages: 1, word: "Papaya")110                let result = try await Fx.run(Fx.request(pdf, to: .pdf, op: PDFEngine.Operation.ocr, ["ocr.pages": .string("all"), "ocr.level": .string("fast")]))111                #expect(result.hasWarning("ocr-duplicate-text") && result.output.facts["ocr.pages"] == "1")112                #expect(Fx.pdfText(result.file, page: 0).components(separatedBy: "Papaya").count - 1 >= 3)113            }114        }115    }116117    @Suite struct RichText {118        static func layoutPDF() -> URL {119            let url = Fx.dir("fixture").appendingPathComponent("layout.pdf")120            let ctx = CGContext(url as CFURL, mediaBox: nil, [kCGPDFContextTitle: "Layout Title"] as CFDictionary)!121            var box = CGRect(x: 0, y: 0, width: 612, height: 792)122            ctx.beginPage(mediaBox: &box)123            Fx.draw("Reconstruction Heading", at: CGPoint(x: 72, y: 700), size: 24, in: ctx)124            Fx.draw("This first paragraph has several lines that wrap across the", at: CGPoint(x: 72, y: 660), size: 12, in: ctx)125            Fx.draw("full width of the column and should be merged into a single", at: CGPoint(x: 72, y: 645), size: 12, in: ctx)126            Fx.draw("paragraph ending here.", at: CGPoint(x: 72, y: 630), size: 12, in: ctx)127            Fx.draw("A second paragraph follows a vertical gap and has a hyphen-", at: CGPoint(x: 72, y: 595), size: 12, in: ctx)128            Fx.draw("ated word plus strong text:", at: CGPoint(x: 72, y: 580), size: 12, in: ctx)129            Fx.draw("Boldword", at: CGPoint(x: 230, y: 580), size: 12, font: "Helvetica-Bold", in: ctx)130            Fx.draw("Slantword", at: CGPoint(x: 300, y: 580), size: 12, font: "Helvetica-Oblique", in: ctx)131            ctx.endPage()132            ctx.beginPage(mediaBox: &box)133            Fx.draw("Second page of the layout fixture.", at: CGPoint(x: 72, y: 700), size: 12, in: ctx)134            ctx.endPage()135            ctx.closePDF()136            return url137        }138139        @Test(arguments: [FormatID.docx, .rtf])140        func rebuildsParagraphsAndStyles(format: FormatID) async throws {141            let result = try await Fx.run(Fx.request(RichText.layoutPDF(), to: format))142            let type: NSAttributedString.DocumentType = format == .rtf ? .rtf : .officeOpenXML143            var attributes: NSDictionary?144            let document = try NSAttributedString(url: result.file, options: [.documentType: type], documentAttributes: &attributes)145            let text = document.string146            let paragraphs = text.components(separatedBy: CharacterSet(charactersIn: "\n\u{0C}\u{2029}")).filter { !$0.isEmpty }147            #expect(paragraphs.contains("Reconstruction Heading"))148            #expect(paragraphs.contains("This first paragraph has several lines that wrap across the full width of the column and should be merged into a single paragraph ending here."))149            #expect(paragraphs.contains { $0.hasPrefix("A second paragraph follows a vertical gap and has a hyphenated word plus strong text:") && $0.contains("Boldword") })150            #expect(text.contains("\u{0C}") && paragraphs.last == "Second page of the layout fixture.")151            func font(of word: String) -> NSFont? {152                let range = (text as NSString).range(of: word)153                return range.location == NSNotFound ? nil : document.attribute(.font, at: range.location, effectiveRange: nil) as? NSFont154            }155            #expect(font(of: "Reconstruction")?.pointSize == 24)156            #expect(font(of: "several")?.pointSize == 12)157            #expect(font(of: "Boldword")?.fontDescriptor.symbolicTraits.contains(.bold) == true)158            #expect(font(of: "Slantword")?.fontDescriptor.symbolicTraits.contains(.italic) == true)159            #expect(font(of: "several")?.fontDescriptor.symbolicTraits.contains(.bold) == false)160            #expect(result.output.facts["headings.detected"] == "1")161            #expect((attributes?[NSAttributedString.DocumentAttributeKey.title] as? String) == "Layout Title")162            if format == .docx { #expect(try Data(contentsOf: result.file).prefix(2) == Data("PK".utf8)) }163            let noBreaks = try await Fx.run(Fx.request(RichText.layoutPDF(), to: format, ["docx.page-breaks": .bool(false), "pages": .string("1")]))164            let single = try NSAttributedString(url: noBreaks.file, options: [.documentType: type], documentAttributes: nil).string165            #expect(!single.contains("\u{0C}") && !single.contains("Second page"))166        }167168        @Test func scannedDocumentsAreRefusedWithAdvice() async throws {169            let pdf = Fx.scannedPDF("scan-docx.pdf", dpi: 72)170            #expect(Fx.engine.preflight(Fx.request(pdf, to: .docx), inspection: nil).contains { $0.code == "scanned-no-text" && $0.severity == .dataLoss })171            do {172                _ = try await Fx.run(Fx.request(pdf, to: .docx))173                Issue.record("a scanned PDF must not convert to DOCX silently")174            } catch let error as ConversionError {175                #expect(error.localizedDescription.contains("OCR"))176            }177        }178    }179}180
Files main
Menu

Theme