Spaces:

NagaYu
/

dasumae

Running

Upload 6 files

c4413a6 verified about 5 hours ago

2.54 kB

// OCR via Tesseract.js, self-hosted, using the EXACT Phase 0 configuration:
//   languages 'jpn+eng', OEM 1 (LSTM_ONLY), PSM 3 (AUTO).
// Tesseract.js runs the wasm in its own Web Worker, so the heavy OCR work stays
// off the main thread and the UI never freezes. All paths point at the bundled
// vendor/ copies — no third-party request is made.

// The vendored ESM build exposes a single default export (the Tesseract object);
// createWorker is a property of it.
import Tesseract from '../vendor/tesseract/tesseract.esm.min.js';
const { createWorker } = Tesseract;

const WORKER_PATH = new URL('../vendor/tesseract/worker.min.js', import.meta.url).href;
const CORE_PATH = new URL('../vendor/tesseract/tesseract-core-simd-lstm.js', import.meta.url).href;
const LANG_PATH = new URL('../vendor/tessdata/', import.meta.url).href;

// Extract word-level boxes exactly as the Phase 0 spike did, so detect.js sees
// the same shape of input it was measured against.
function wordsFromData(data) {
  const words = [];
  for (const b of data.blocks || []) {
    for (const p of b.paragraphs || []) {
      for (const l of p.lines || []) {
        for (const w of l.words || []) {
          if (!w.text || !w.text.trim()) continue;
          const { x0, y0, x1, y1 } = w.bbox;
          words.push({ text: w.text, confidence: w.confidence, bbox: { x: x0, y: y0, w: x1 - x0, h: y1 - y0 } });
        }
      }
    }
  }
  return words;
}

let workerPromise = null;

export function initOcr(onProgress) {
  if (workerPromise) return workerPromise;
  workerPromise = (async () => {
    const worker = await createWorker('jpn+eng', 1, {
      workerPath: WORKER_PATH,
      corePath: CORE_PATH,
      langPath: LANG_PATH,
      // Run the worker from its real file URL (not a blob:) so the emscripten
      // core resolves its sibling .wasm relative to vendor/tesseract/.
      workerBlobURL: false,
      cacheMethod: 'none', // always load from the bundled vendor copy (offline-safe)
      gzip: true,
      logger: (m) => { if (onProgress) onProgress(m); },
    });
    await worker.setParameters({ tessedit_pageseg_mode: '3' }); // PSM 3 AUTO
    return worker;
  })();
  return workerPromise;
}

// Run OCR on a drawable source (canvas at natural size) -> { words, ms }.
export async function runOcr(source, onProgress) {
  const worker = await initOcr(onProgress);
  const t0 = performance.now();
  const { data } = await worker.recognize(source, {}, { blocks: true });
  return { words: wordsFromData(data), ms: performance.now() - t0 };
}

	// OCR via Tesseract.js, self-hosted, using the EXACT Phase 0 configuration:
	// languages 'jpn+eng', OEM 1 (LSTM_ONLY), PSM 3 (AUTO).
	// Tesseract.js runs the wasm in its own Web Worker, so the heavy OCR work stays
	// off the main thread and the UI never freezes. All paths point at the bundled
	// vendor/ copies — no third-party request is made.

	// The vendored ESM build exposes a single default export (the Tesseract object);
	// createWorker is a property of it.
	import Tesseract from '../vendor/tesseract/tesseract.esm.min.js';
	const { createWorker } = Tesseract;

	const WORKER_PATH = new URL('../vendor/tesseract/worker.min.js', import.meta.url).href;
	const CORE_PATH = new URL('../vendor/tesseract/tesseract-core-simd-lstm.js', import.meta.url).href;
	const LANG_PATH = new URL('../vendor/tessdata/', import.meta.url).href;

	// Extract word-level boxes exactly as the Phase 0 spike did, so detect.js sees
	// the same shape of input it was measured against.
	function wordsFromData(data) {
	const words = [];
	for (const b of data.blocks \|\| []) {
	for (const p of b.paragraphs \|\| []) {
	for (const l of p.lines \|\| []) {
	for (const w of l.words \|\| []) {
	if (!w.text \|\| !w.text.trim()) continue;
	const { x0, y0, x1, y1 } = w.bbox;
	words.push({ text: w.text, confidence: w.confidence, bbox: { x: x0, y: y0, w: x1 - x0, h: y1 - y0 } });
	}
	}
	}
	}
	return words;
	}

	let workerPromise = null;

	export function initOcr(onProgress) {
	if (workerPromise) return workerPromise;
	workerPromise = (async () => {
	const worker = await createWorker('jpn+eng', 1, {
	workerPath: WORKER_PATH,
	corePath: CORE_PATH,
	langPath: LANG_PATH,
	// Run the worker from its real file URL (not a blob:) so the emscripten
	// core resolves its sibling .wasm relative to vendor/tesseract/.
	workerBlobURL: false,
	cacheMethod: 'none', // always load from the bundled vendor copy (offline-safe)
	gzip: true,
	logger: (m) => { if (onProgress) onProgress(m); },
	});
	await worker.setParameters({ tessedit_pageseg_mode: '3' }); // PSM 3 AUTO
	return worker;
	})();
	return workerPromise;
	}

	// Run OCR on a drawable source (canvas at natural size) -> { words, ms }.
	export async function runOcr(source, onProgress) {
	const worker = await initOcr(onProgress);
	const t0 = performance.now();
	const { data } = await worker.recognize(source, {}, { blocks: true });
	return { words: wordsFromData(data), ms: performance.now() - t0 };
	}