472 lines
19 KiB
JavaScript
472 lines
19 KiB
JavaScript
// js/process/ocr-adapters/local-adapter.js
|
||
// 本地 PDF 解析适配器 - 使用 PDF.js 直接提取 PDF 文本
|
||
|
||
/**
|
||
* 本地 PDF 解析适配器
|
||
* 特点：
|
||
* - 免费、快速、无需 API Key
|
||
* - 仅适用于文字型 PDF（非扫描件）
|
||
* - 不支持 OCR（扫描图片转文字）
|
||
* - 尽力保留图片和表格的位置顺序
|
||
*/
|
||
class LocalPdfAdapter extends BaseOcrAdapter {
|
||
constructor() {
|
||
super();
|
||
this.name = 'LocalPDF';
|
||
|
||
// 检查 PDF.js 是否加载
|
||
if (typeof pdfjsLib === 'undefined') {
|
||
throw new Error('PDF.js 库未加载，无法使用本地 PDF 解析功能');
|
||
}
|
||
}
|
||
|
||
/**
|
||
* 验证配置
|
||
* 本地解析不需要任何配置
|
||
*/
|
||
async validateConfig() {
|
||
return { valid: true };
|
||
}
|
||
|
||
/**
|
||
* 处理 PDF 文件
|
||
* @param {File} file - PDF 文件对象
|
||
* @param {Function} onProgress - 进度回调 (current, total, message)
|
||
* @returns {Promise<{markdown: string, images: Array}>}
|
||
*/
|
||
async processFile(file, onProgress) {
|
||
try {
|
||
onProgress(0, 100, '开始本地解析 PDF...');
|
||
|
||
// 读取文件为 ArrayBuffer
|
||
const arrayBuffer = await file.arrayBuffer();
|
||
|
||
onProgress(10, 100, '加载 PDF 文档...');
|
||
|
||
// 加载 PDF 文档
|
||
const loadingTask = pdfjsLib.getDocument({ data: arrayBuffer });
|
||
const pdf = await loadingTask.promise;
|
||
|
||
const numPages = pdf.numPages;
|
||
onProgress(20, 100, `PDF 共 ${numPages} 页，开始提取内容...`);
|
||
|
||
const images = [];
|
||
const pageContents = []; // 存储每页的内容片段（文本 + 图片标记）
|
||
const BATCH_SIZE = 50; // 每50页处理一批，避免内存累积
|
||
const totalBatches = Math.ceil(numPages / BATCH_SIZE);
|
||
|
||
console.log(`[LocalPDF] 将分 ${totalBatches} 批处理，每批 ${BATCH_SIZE} 页`);
|
||
|
||
// 分批处理页面
|
||
for (let batchIdx = 0; batchIdx < totalBatches; batchIdx++) {
|
||
const startPage = batchIdx * BATCH_SIZE + 1;
|
||
const endPage = Math.min((batchIdx + 1) * BATCH_SIZE, numPages);
|
||
|
||
console.log(`[LocalPDF] 处理第 ${batchIdx + 1}/${totalBatches} 批：第 ${startPage}-${endPage} 页`);
|
||
|
||
// 处理当前批次的页面
|
||
await this._processBatch(pdf, startPage, endPage, numPages, images, pageContents, onProgress);
|
||
|
||
// 批次间休息，让浏览器释放内存
|
||
if (batchIdx < totalBatches - 1) {
|
||
console.log(`[LocalPDF] 第 ${batchIdx + 1} 批完成，休息500ms...`);
|
||
await this.sleep(500);
|
||
}
|
||
}
|
||
|
||
onProgress(90, 100, '启发式文本重建中...');
|
||
|
||
// 合并所有页面内容（去掉页间分隔符）
|
||
const mergedText = pageContents.join('\n\n');
|
||
|
||
// 启发式文本重建
|
||
const rebuiltText = this._heuristicRebuild(mergedText);
|
||
|
||
onProgress(100, 100, '本地解析完成');
|
||
|
||
return {
|
||
markdown: rebuiltText,
|
||
images: images,
|
||
metadata: {
|
||
engine: 'local',
|
||
source: 'pdfjs',
|
||
pages: numPages,
|
||
note: '本地解析提取文本并保留图片位置'
|
||
}
|
||
};
|
||
} catch (error) {
|
||
console.error('[LocalPDF] 处理失败:', error);
|
||
throw new Error(`本地 PDF 解析失败: ${error.message}`);
|
||
}
|
||
}
|
||
|
||
/**
|
||
* 处理一批页面
|
||
* @private
|
||
*/
|
||
async _processBatch(pdf, startPage, endPage, totalPages, images, pageContents, onProgress) {
|
||
for (let pageNum = startPage; pageNum <= endPage; pageNum++) {
|
||
try {
|
||
onProgress(
|
||
20 + Math.floor((pageNum / totalPages) * 70),
|
||
100,
|
||
`正在解析第 ${pageNum}/${totalPages} 页...`
|
||
);
|
||
|
||
console.log(`[LocalPDF] 开始处理第 ${pageNum} 页...`);
|
||
const page = await pdf.getPage(pageNum);
|
||
|
||
console.log(`[LocalPDF] 第 ${pageNum} 页：提取文本...`);
|
||
const textContent = await page.getTextContent();
|
||
const pageText = this._extractTextFromPage(textContent);
|
||
|
||
console.log(`[LocalPDF] 第 ${pageNum} 页：检测图片...`);
|
||
// 检测图片位置并插入占位符（带超时保护）
|
||
const pageContentWithImages = await Promise.race([
|
||
this._insertImagePlaceholders(page, pageText, pageNum, images),
|
||
new Promise((_, reject) =>
|
||
setTimeout(() => reject(new Error('图片提取超时')), 30000) // 30秒超时
|
||
)
|
||
]).catch(err => {
|
||
console.warn(`[LocalPDF] 第 ${pageNum} 页图片提取失败:`, err);
|
||
return pageText; // 失败时只返回文本
|
||
});
|
||
|
||
pageContents.push(pageContentWithImages);
|
||
|
||
// 清理页面资源，避免内存泄漏
|
||
page.cleanup();
|
||
|
||
// 每10页输出一次进度日志
|
||
if (pageNum % 10 === 0) {
|
||
console.log(`[LocalPDF] ✓ 已处理 ${pageNum}/${totalPages} 页，已提取 ${images.length} 个图片`);
|
||
// 每10页强制等待一下，让浏览器释放内存
|
||
await this.sleep(200);
|
||
}
|
||
} catch (pageError) {
|
||
console.error(`[LocalPDF] ✗ 第 ${pageNum} 页处理失败:`, pageError);
|
||
// 页面失败不中断整体处理，记录错误并继续
|
||
pageContents.push(`\n\n*[第 ${pageNum} 页解析失败: ${pageError.message}]*\n\n`);
|
||
}
|
||
}
|
||
}
|
||
|
||
/**
|
||
* 在原页面位置插入图片占位符
|
||
* @private
|
||
*/
|
||
async _insertImagePlaceholders(page, pageText, pageNum, images) {
|
||
try {
|
||
const ops = await page.getOperatorList();
|
||
|
||
// 收集图片对象名称和位置
|
||
const imageInfos = [];
|
||
for (let i = 0; i < ops.fnArray.length; i++) {
|
||
const fn = ops.fnArray[i];
|
||
if (fn === pdfjsLib.OPS.paintImageXObject || fn === pdfjsLib.OPS.paintJpegXObject) {
|
||
const imgName = ops.argsArray[i][0];
|
||
imageInfos.push({ name: imgName, opIndex: i });
|
||
}
|
||
}
|
||
|
||
if (imageInfos.length === 0) {
|
||
return pageText;
|
||
}
|
||
|
||
console.log(`[LocalPDF] 第 ${pageNum} 页检测到 ${imageInfos.length} 个图片对象`);
|
||
let result = pageText;
|
||
let extractedCount = 0;
|
||
let skippedCount = 0;
|
||
|
||
// 尝试提取每个图片
|
||
for (let idx = 0; idx < imageInfos.length; idx++) {
|
||
const imageId = `page${pageNum}_img${idx + 1}`;
|
||
const imgInfo = imageInfos[idx];
|
||
|
||
try {
|
||
// 给每个图片的获取操作添加超时
|
||
// 使用更长的超时时间，但只对前几个图片尝试
|
||
const timeout = (idx < 3) ? 2000 : 1000; // 前3个图片等2秒，其他等1秒
|
||
const imgObj = await Promise.race([
|
||
new Promise((resolve) => {
|
||
page.objs.get(imgInfo.name, resolve);
|
||
}),
|
||
new Promise((_, reject) =>
|
||
setTimeout(() => reject(new Error('获取图片对象超时')), timeout)
|
||
)
|
||
]);
|
||
|
||
if (!imgObj) {
|
||
console.warn(`[LocalPDF] 图片对象 ${imageId} 为空，跳过`);
|
||
skippedCount++;
|
||
continue;
|
||
}
|
||
|
||
const width = imgObj.width;
|
||
const height = imgObj.height;
|
||
|
||
// 过滤规则：
|
||
// 1. 跳过超小图（装饰）：宽或高 < 30px
|
||
// 2. 保留合理宽高比：0.05 - 20
|
||
const aspectRatio = width / height;
|
||
const isTooSmall = width < 30 || height < 30;
|
||
const isBadAspectRatio = aspectRatio < 0.05 || aspectRatio > 20;
|
||
|
||
if (isTooSmall) {
|
||
console.log(`[LocalPDF] 跳过超小图片 ${imageId}: ${width}x${height} (可能是图标或装饰)`);
|
||
skippedCount++;
|
||
continue;
|
||
}
|
||
|
||
if (isBadAspectRatio) {
|
||
console.log(`[LocalPDF] 跳过异常宽高比图片 ${imageId}: ${width}x${height} (ratio: ${aspectRatio.toFixed(2)})`);
|
||
skippedCount++;
|
||
continue;
|
||
}
|
||
|
||
console.log(`[LocalPDF] 提取图片 ${imageId}: ${width}x${height}`);
|
||
let base64Data = null;
|
||
|
||
// 方法1: ImageBitmap 对象（PDF.js 常见格式）
|
||
if (imgObj.bitmap && imgObj.bitmap instanceof ImageBitmap) {
|
||
// 限制图片最大尺寸，进一步降低以减少内存
|
||
const maxDimension = 600; // 降低到600px
|
||
let targetWidth = width;
|
||
let targetHeight = height;
|
||
|
||
if (targetWidth > maxDimension || targetHeight > maxDimension) {
|
||
const scale = Math.min(maxDimension / targetWidth, maxDimension / targetHeight);
|
||
targetWidth = Math.floor(targetWidth * scale);
|
||
targetHeight = Math.floor(targetHeight * scale);
|
||
}
|
||
|
||
const canvas = document.createElement('canvas');
|
||
canvas.width = targetWidth;
|
||
canvas.height = targetHeight;
|
||
const ctx = canvas.getContext('2d', { willReadFrequently: false });
|
||
|
||
// 使用 drawImage 绘制并缩放 ImageBitmap
|
||
ctx.drawImage(imgObj.bitmap, 0, 0, targetWidth, targetHeight);
|
||
base64Data = canvas.toDataURL('image/jpeg', 0.7); // 降低质量到70%
|
||
|
||
// 立即释放资源
|
||
ctx.clearRect(0, 0, canvas.width, canvas.height);
|
||
canvas.width = 0;
|
||
canvas.height = 0;
|
||
|
||
// 关闭 ImageBitmap
|
||
if (imgObj.bitmap.close) {
|
||
imgObj.bitmap.close();
|
||
}
|
||
}
|
||
// 方法2: JPEG 原始数据
|
||
else if (imgObj.kind === 1 && imgObj.data) {
|
||
const bytes = imgObj.data;
|
||
const blob = new Blob([bytes], { type: 'image/jpeg' });
|
||
base64Data = await this.blobToBase64(blob);
|
||
}
|
||
// 方法3: 原始像素数据
|
||
else if (imgObj.data) {
|
||
const imgData = imgObj.data;
|
||
|
||
const canvas = document.createElement('canvas');
|
||
canvas.width = width;
|
||
canvas.height = height;
|
||
const ctx = canvas.getContext('2d', { willReadFrequently: false });
|
||
|
||
const imageData = ctx.createImageData(width, height);
|
||
|
||
if (imgData instanceof Uint8ClampedArray || imgData instanceof Uint8Array) {
|
||
imageData.data.set(imgData);
|
||
} else if (ArrayBuffer.isView(imgData)) {
|
||
imageData.data.set(new Uint8ClampedArray(imgData.buffer));
|
||
} else {
|
||
throw new Error(`Unsupported image data format: ${imgData.constructor.name}`);
|
||
}
|
||
|
||
ctx.putImageData(imageData, 0, 0);
|
||
base64Data = canvas.toDataURL('image/jpeg', 0.7); // 降低质量到70%
|
||
|
||
canvas.width = 0;
|
||
canvas.height = 0;
|
||
} else {
|
||
throw new Error('Unknown image format');
|
||
}
|
||
|
||
if (base64Data) {
|
||
images.push({
|
||
id: imageId,
|
||
data: base64Data
|
||
});
|
||
|
||
// 在文本末尾添加图片引用
|
||
result += `\n\n\n\n`;
|
||
extractedCount++;
|
||
}
|
||
} catch (imgError) {
|
||
// 图片提取失败，静默跳过，只在需要调试时输出
|
||
if (imgError.message.includes('超时')) {
|
||
// 超时错误不输出，太多了
|
||
skippedCount++;
|
||
} else {
|
||
console.warn(`[LocalPDF] 提取图片 ${imageId} 失败:`, imgError.message);
|
||
skippedCount++;
|
||
}
|
||
}
|
||
|
||
// 每处理5个图片休息一下，让浏览器有时间释放内存
|
||
if ((idx + 1) % 5 === 0) {
|
||
await this.sleep(100);
|
||
}
|
||
}
|
||
|
||
console.log(`[LocalPDF] 第 ${pageNum} 页图片处理完成: 提取 ${extractedCount} 个，跳过 ${skippedCount} 个`);
|
||
return result;
|
||
} catch (error) {
|
||
console.warn(`[LocalPDF] 第 ${pageNum} 页图片处理失败:`, error);
|
||
return pageText;
|
||
}
|
||
}
|
||
|
||
/**
|
||
* 启发式文本重建
|
||
* @private
|
||
*/
|
||
_heuristicRebuild(text) {
|
||
let rebuilt = text;
|
||
|
||
// 先保护图片引用，避免被文本处理规则破坏
|
||
const imageRefs = [];
|
||
rebuilt = rebuilt.replace(/!\[([^\]]*)\]\(([^)]+)\)/g, (match) => {
|
||
const placeholder = `__IMG_PLACEHOLDER_${imageRefs.length}__`;
|
||
imageRefs.push(match);
|
||
return placeholder;
|
||
});
|
||
|
||
// 1. 修复被断开的单词（英文）
|
||
// 匹配：字母-空格-换行-字母 -> 字母字母
|
||
rebuilt = rebuilt.replace(/([a-zA-Z])-\s*\n\s*([a-z])/g, '$1$2');
|
||
|
||
// 2. 合并被打断的句子
|
||
// 如果行尾不是句号等结束符，且下一行不是大写/数字/特殊字符开头，则合并
|
||
rebuilt = rebuilt.replace(/([^\n.!?。！？])\n([a-z\u4e00-\u9fa5])/g, '$1 $2');
|
||
|
||
// 3. 修复标点符号周围的空格
|
||
// 中文标点前后不应有空格
|
||
rebuilt = rebuilt.replace(/\s+([，。！？；：、）】」』])/g, '$1');
|
||
rebuilt = rebuilt.replace(/([（【「『])\s+/g, '$1');
|
||
|
||
// 英文标点后应有空格（如果后面是字母）
|
||
rebuilt = rebuilt.replace(/([,.!?;:])([a-zA-Z])/g, '$1 $2');
|
||
|
||
// 移除标点前的多余空格
|
||
rebuilt = rebuilt.replace(/\s+([,.!?;:])/g, '$1');
|
||
|
||
// 4. 规范化空白字符
|
||
// 多个空格变成一个
|
||
rebuilt = rebuilt.replace(/ {2,}/g, ' ');
|
||
|
||
// 保留段落分隔（最多2个换行）
|
||
rebuilt = rebuilt.replace(/\n{3,}/g, '\n\n');
|
||
|
||
// 5. 修复常见的格式问题
|
||
// 修复：数字. 后面应该有空格（列表项）
|
||
rebuilt = rebuilt.replace(/(\d+)\.\s*([a-zA-Z\u4e00-\u9fa5])/g, '$1. $2');
|
||
|
||
// 修复：括号内不应有首尾空格
|
||
rebuilt = rebuilt.replace(/\(\s+/g, '(');
|
||
rebuilt = rebuilt.replace(/\s+\)/g, ')');
|
||
|
||
// 6. 智能段落识别
|
||
// 如果连续的短行可能是同一段落，尝试合并
|
||
const lines = rebuilt.split('\n');
|
||
const paragraphs = [];
|
||
let currentPara = '';
|
||
|
||
for (let i = 0; i < lines.length; i++) {
|
||
const line = lines[i].trim();
|
||
|
||
if (line === '') {
|
||
if (currentPara) {
|
||
paragraphs.push(currentPara.trim());
|
||
currentPara = '';
|
||
}
|
||
continue;
|
||
}
|
||
|
||
// 判断是否应该换段
|
||
const shouldBreak =
|
||
// 当前段落为空
|
||
currentPara === '' ||
|
||
// 行以标题标记开头
|
||
/^#{1,6}\s/.test(line) ||
|
||
// 行以列表标记开头
|
||
/^[\-\*\+]\s/.test(line) ||
|
||
/^\d+\.\s/.test(line) ||
|
||
// 上一行以句号等结束且本行首字母大写
|
||
(/[.!?。！？]\s*$/.test(currentPara) && /^[A-Z\u4e00-\u9fa5]/.test(line));
|
||
|
||
if (shouldBreak) {
|
||
if (currentPara) {
|
||
paragraphs.push(currentPara.trim());
|
||
}
|
||
currentPara = line;
|
||
} else {
|
||
currentPara += ' ' + line;
|
||
}
|
||
}
|
||
|
||
if (currentPara) {
|
||
paragraphs.push(currentPara.trim());
|
||
}
|
||
|
||
rebuilt = paragraphs.join('\n\n');
|
||
|
||
// 恢复图片引用
|
||
imageRefs.forEach((ref, idx) => {
|
||
rebuilt = rebuilt.replace(`__IMG_PLACEHOLDER_${idx}__`, ref);
|
||
});
|
||
|
||
return rebuilt.trim();
|
||
}
|
||
|
||
/**
|
||
* 从 TextContent 中提取文本
|
||
* @private
|
||
*/
|
||
_extractTextFromPage(textContent) {
|
||
const items = textContent.items;
|
||
let text = '';
|
||
let lastY = null;
|
||
|
||
for (let i = 0; i < items.length; i++) {
|
||
const item = items[i];
|
||
|
||
// 检测换行（Y 坐标变化）
|
||
if (lastY !== null && Math.abs(item.transform[5] - lastY) > 5) {
|
||
text += '\n';
|
||
}
|
||
|
||
text += item.str;
|
||
|
||
// 如果下一个 item 有空格，添加空格
|
||
if (i < items.length - 1) {
|
||
const nextItem = items[i + 1];
|
||
const spaceWidth = item.width * 0.3; // 估算空格宽度
|
||
if (nextItem.transform[4] - item.transform[4] > spaceWidth) {
|
||
text += ' ';
|
||
}
|
||
}
|
||
|
||
lastY = item.transform[5];
|
||
}
|
||
|
||
return text.trim();
|
||
}
|
||
}
|
||
|
||
// 注册到全局
|
||
if (typeof window !== 'undefined') {
|
||
window.LocalPdfAdapter = LocalPdfAdapter;
|
||
}
|