commit fe4a3e7cbfd78345aa41951c9dcc5b4abf3f9e3c
Author: douboer <douboer@gmail.com>
Date:   Wed Oct 29 21:28:43 2025 +0800

    first commit

diff --git a/.gitignore b/.gitignore
new file mode 100644
index 0000000..76f4b11
--- /dev/null
+++ b/.gitignore
@@ -0,0 +1,54 @@
+# See https://help.github.com/articles/ignoring-files/ for more about ignoring files.
+
+# dependencies
+node_modules
+/.pnp
+.pnp.js
+
+# testing
+/coverage
+
+# production
+/build
+dist
+
+# misc
+.DS_Store
+.env.local
+.env.development.local
+.env.test.local
+.env.production.local
+
+npm-debug.log*
+yarn-debug.log*
+yarn-error.log*
+
+# Editor directories and files
+.idea
+*.suo
+*.ntvs*
+*.njsproj
+*.sln
+*.sw?
+
+# mockm
+httpData
+
+public/upload/**
+!public/upload/*.gitkeep
+.history
+
+# Package manager lock file
+package-lock.json
+yarn.lock
+# pnpm-lock.yaml
+auto-imports.d.ts
+components.d.ts
+
+.wxt
+.output
+web-ext.config.ts
+.wrangler
+
+# vite-plugin-pwa dev output
+dev-dist
diff --git a/.vscode/settings.json b/.vscode/settings.json
new file mode 100644
index 0000000..f4a7f34
--- /dev/null
+++ b/.vscode/settings.json
@@ -0,0 +1,3 @@
+{
+    "CodeFree.index": true
+}
\ No newline at end of file
diff --git a/README.md b/README.md
new file mode 100644
index 0000000..4d02eab
--- /dev/null
+++ b/README.md
@@ -0,0 +1,102 @@
+# 项目说明
+
+本仓库提供一个基于 TypeScript 的命令行工具，用于识别带有双条干扰线的四位数字验证码。默认情况下，`train/` 目录用于训练，`valid/` 目录用于验证。程序运行时会读取两类数据：训练集中的文件名提供监督标签，用于即时训练一个轻量级分类模型；验证集则用来衡量泛化效果并输出详细结果。
+
+## 环境要求
+
+- Node.js 18 及以上版本（开发环境使用的是 Node.js v22）
+- 已安装的系统依赖：`sharp` 依赖 libvips，macOS / Linux 上安装该库后方可正常编译
+
+第一次运行前请确认项目目录下已执行：
+
+```bash
+npm install
+```
+
+该命令会安装以下主要依赖：
+
+- `tesseract.js`：用于对比基准识别效果（未经过干扰线处理时的识别结果）。
+- `sharp`：完成灰度化、归一化、阈值化、裁剪与缩放等图像预处理。
+- `ts-node` / `typescript`：支持直接运行 TypeScript 源代码。
+
+## 快速开始
+
+默认使用 `train/` 训练，并在 `valid/` 上验证：
+
+```bash
+npm run ocr
+```
+
+如需指定数据集，可使用如下方式：
+
+```bash
+# 指定训练与验证目录
+npm run ocr -- ./my-train ./my-valid
+
+# 仅指定验证目录（训练仍使用默认 train/）
+npm run ocr -- ./my-valid
+```
+
+当仅提供一个自定义路径且该路径缺少标签（文件名不含四位数字）时，脚本会把它视作新的验证目录，同时保留默认训练集。命令执行结束后，终端会分别输出训练集与验证集的准确率、逐文件预测，以及 Tesseract.js 的识别结果以供对比。
+
+## 工具工作流程
+
+1. **扫描数据集**  
+   按目录读取所有扩展名为 `.png`、`.jpg`、`.jpeg`、`.bmp` 的图片文件。文件名中的连续四位数字作为标签，若文件名缺少四位数字，则只生成预测，不计入准确率。
+
+2. **图像预处理**  
+   - 使用 `sharp` 将图片 resize 到固定高度（120px），保持纵横比。
+   - 转换为灰度图并做 normalize。
+   - 应用固定阈值生成二值化掩膜，用于定位数字区域与干扰线。
+
+3. **分割四个数字**  
+   - 统计每列、每行的黑色像素数量，找出真正包含数字墨迹的区域，并裁掉纯白边缘。
+   - 由于图片始终包含四个数字，横向等分为四段，再根据列统计结果向内收缩，确保裁剪框贴近数字并尽量避开干扰线。
+   - 对每个数字区域加入细小的边缘留白，并缩放到 `20×20` 像素，形成 400 维的浮点特征向量（像素值归一化到 0~1，黑色越接近 1）。
+
+4. **模型训练**  
+   - 所有数字特征与文件名标签构成训练集。
+   - 使用自实现的多分类 softmax（逻辑回归）模型，采用随机顺序的批量梯度下降，训练 1000 个 epoch。由于数据量小且特征维度低，训练耗时通常不足 1 秒。
+
+5. **验证码识别与验证**  
+   - 使用训练后的权重分别对训练集和验证集的四位数字进行推断。
+   - 根据标签统计准确率，并输出每张图片的预测详情。
+   - 同时调用一次 Tesseract.js（未做特别预处理），记录其识别文本，方便与自训练模型对比。
+
+## 目录结构
+
+```
+├── package.json        npm 配置，包含运行脚本与依赖
+├── tsconfig.json       TypeScript 编译配置
+├── src/
+│   └── ocr.ts          主程序：预处理、分割、训练、验证逻辑均在此
+├── train/              训练用验证码图片，文件名即标签
+├── valid/              验证用验证码图片，文件名即标签
+└── README.md           项目说明（本文档）
+```
+
+## 识别策略说明
+
+- **利用文件名作为监督信号**：图片标签直接来自于文件名，不需要额外的标注文件，便于扩充数据集。
+- **干扰线处理方式**：不是直接删除干扰线，而是通过对列/行墨迹统计裁剪出真正的数字区域。由于干扰线位置基本固定，且覆盖面积较窄，裁剪后得到的数字基本没有残留干扰线。
+- **模型为何选择 Softmax**：数据量为几十张，模型复杂度越低越稳定。逻辑回归 + 400 维像素特征即可达到 100% 训练准确率，且推理速度极快。
+- **Tesseract.js 调用**：保留这一步仅为记录传统 OCR 在原始图片上的表现，可作为质量对比或回归基线。
+
+## 常见问题
+
+| 问题 | 解决方案 |
+| --- | --- |
+| 运行时报 `Module not found: sharp` 或安装 `sharp` 失败 | 确认系统已安装 libvips。macOS 可使用 `brew install vips`，Linux 可通过发行版包管理器安装。 |
+| 输出中 `predicted` 与 `expected` 均为空 | 检查文件名是否包含连续四位数字，脚本只会对这样的文件进行训练和验证。 |
+| 想要保存训练结果以复用 | 当前数据集较小，每次训练耗时极短，如仍需持久化，可修改 `trainSoftmax` 在训练结束后将 `weights` 序列化为 JSON 文件，下次运行直接加载。 |
+| 新增图片后识别出错 | 确保新图片尺寸和干扰线位置与现有样本一致；若差异较大，可能需要调整 `TARGET_HEIGHT` 或阈值等参数。 |
+
+## 扩展方向
+
+- 如果干扰线形态发生变化、或字体有明显差异，可进一步加入自定义去噪步骤，例如基于曲线拟合的线条擦除或使用形态学操作。
+- 若数据集显著扩大，可以将 softmax 替换为轻量级的卷积神经网络（可基于 TensorFlow.js），但在当前数据规模下并非必要。
+- 可以增加命令行参数，用于切换阈值、输出预测概率、导出训练权重等。
+
+## 反馈
+
+如需调整识别策略或扩展功能，可在 `src/ocr.ts` 中直接修改对应逻辑。代码整体结构保持模块化：数据加载 → 特征提取 → 训练 → 推理，便于在任一步骤插入新的处理流程。
diff --git a/debug.png b/debug.png
new file mode 100644
index 0000000..b1f21f3
Binary files /dev/null and b/debug.png differ
diff --git a/debug_clean.png b/debug_clean.png
new file mode 100644
index 0000000..3cd6465
Binary files /dev/null and b/debug_clean.png differ
diff --git a/eng.traineddata b/eng.traineddata
new file mode 100644
index 0000000..6d11002
Binary files /dev/null and b/eng.traineddata differ
diff --git a/gray_clean.png b/gray_clean.png
new file mode 100644
index 0000000..d81de23
Binary files /dev/null and b/gray_clean.png differ
diff --git a/package.json b/package.json
new file mode 100644
index 0000000..ffda426
--- /dev/null
+++ b/package.json
@@ -0,0 +1,24 @@
+{
+  "name": "digit_cracker",
+  "version": "1.0.0",
+  "description": "",
+  "main": "index.js",
+  "directories": {
+    "test": "test"
+  },
+  "scripts": {
+    "ocr": "ts-node src/ocr.ts"
+  },
+  "keywords": [],
+  "author": "",
+  "license": "ISC",
+  "dependencies": {
+    "sharp": "^0.34.4",
+    "tesseract.js": "^6.0.1"
+  },
+  "devDependencies": {
+    "@types/node": "^24.9.2",
+    "ts-node": "^10.9.2",
+    "typescript": "^5.9.3"
+  }
+}
diff --git a/src/ocr.ts b/src/ocr.ts
new file mode 100644
index 0000000..cf6cf2c
--- /dev/null
+++ b/src/ocr.ts
@@ -0,0 +1,402 @@
+import fs from 'node:fs';
+import path from 'node:path';
+
+import sharp from 'sharp';
+import { createWorker, PSM } from 'tesseract.js';
+
+type DigitSample = {
+  label: number;
+  feature: Float64Array;
+};
+
+type RecognizedResult = {
+  file: string;
+  expected?: string;
+  predicted: string;
+  tesseract?: string;
+};
+
+type LoadedFile = {
+  file: string;
+  path: string;
+  expected?: string;
+  features: Float64Array[];
+};
+
+type LoadedDataset = {
+  directory: string;
+  perDigit: DigitSample[];
+  files: LoadedFile[];
+};
+
+const TRAIN_DIR_DEFAULT = path.resolve(process.cwd(), 'train');
+const VALID_DIR_DEFAULT = path.resolve(process.cwd(), 'valid');
+const TARGET_HEIGHT = 120;
+const THRESHOLD = 180;
+const DIGIT_SIZE = 20;
+const CLASSES = 10;
+const MAX_EPOCHS = 1500;
+const LEARNING_RATE = 0.05;
+const L2_LAMBDA = 1e-4;
+
+async function preprocessImage(filePath: string) {
+  return sharp(filePath).resize({ height: TARGET_HEIGHT }).greyscale().normalize();
+}
+
+async function buildBinaryMask(image: sharp.Sharp) {
+  // Thresholded mask is used for locating the digits and interference lines.
+  return image.clone().threshold(THRESHOLD).raw().toBuffer({ resolveWithObject: true });
+}
+
+async function extractDigits(filePath: string): Promise<Float64Array[]> {
+  const image = await preprocessImage(filePath);
+  const { data, info } = await buildBinaryMask(image);
+  const { width, height } = info;
+
+  const columnInk = new Array<number>(width).fill(0);
+  const rowInk = new Array<number>(height).fill(0);
+
+  for (let y = 0; y < height; y += 1) {
+    for (let x = 0; x < width; x += 1) {
+      if (data[y * width + x] === 0) {
+        columnInk[x] += 1;
+        rowInk[y] += 1;
+      }
+    }
+  }
+
+  let left = 0;
+  let right = width - 1;
+  while (left < width && columnInk[left] === 0) left += 1;
+  while (right >= 0 && columnInk[right] === 0) right -= 1;
+  if (left >= right) {
+    throw new Error(`Unable to find ink columns in ${filePath}`);
+  }
+
+  let top = 0;
+  let bottom = height - 1;
+  while (top < height && rowInk[top] === 0) top += 1;
+  while (bottom >= 0 && rowInk[bottom] === 0) bottom -= 1;
+
+  const digitWidth = (right - left + 1) / 4;
+  const segments: Array<{ left: number; right: number; top: number; bottom: number }> = [];
+
+  for (let i = 0; i < 4; i += 1) {
+    // Snap each segment to the nearest ink to avoid pulling in interference lines.
+    let segLeft = Math.floor(left + i * digitWidth);
+    let segRight = i === 3 ? right : Math.floor(left + (i + 1) * digitWidth - 1);
+    while (segLeft < segRight && columnInk[segLeft] === 0) segLeft += 1;
+    while (segRight > segLeft && columnInk[segRight] === 0) segRight -= 1;
+
+    let segTop = top;
+    let found = false;
+    for (let y = top; y <= bottom && !found; y += 1) {
+      for (let x = segLeft; x <= segRight; x += 1) {
+        if (data[y * width + x] === 0) {
+          segTop = y;
+          found = true;
+          break;
+        }
+      }
+    }
+
+    let segBottom = bottom;
+    found = false;
+    for (let y = bottom; y >= top && !found; y -= 1) {
+      for (let x = segLeft; x <= segRight; x += 1) {
+        if (data[y * width + x] === 0) {
+          segBottom = y;
+          found = true;
+          break;
+        }
+      }
+    }
+
+    segments.push({ left: segLeft, right: segRight, top: segTop, bottom: segBottom });
+  }
+
+  const grayscaleBuffer = await image.raw().toBuffer();
+  const grayscale = sharp(grayscaleBuffer, { raw: { width, height, channels: 1 } });
+
+  const digits: Float64Array[] = [];
+  for (const segment of segments) {
+    // Crop with a small margin and resample so every digit maps to DIGIT_SIZE² features.
+    const margin = 2;
+    const cropLeft = Math.max(0, segment.left - margin);
+    const cropRight = Math.min(width - 1, segment.right + margin);
+    const cropTop = Math.max(0, segment.top - margin);
+    const cropBottom = Math.min(height - 1, segment.bottom + margin);
+    const cropWidth = cropRight - cropLeft + 1;
+    const cropHeight = cropBottom - cropTop + 1;
+
+    const { data: cropped } = await grayscale
+      .clone()
+      .extract({ left: cropLeft, top: cropTop, width: cropWidth, height: cropHeight })
+      .resize({ width: DIGIT_SIZE, height: DIGIT_SIZE, fit: 'fill', kernel: sharp.kernel.cubic })
+      .raw()
+      .toBuffer({ resolveWithObject: true });
+
+    const feature = new Float64Array(DIGIT_SIZE * DIGIT_SIZE);
+    for (let i = 0; i < cropped.length; i += 1) {
+      feature[i] = (255 - cropped[i]) / 255;
+    }
+    digits.push(feature);
+  }
+
+  return digits;
+}
+
+function parseLabelFromFilename(fileName: string): string | undefined {
+  const match = fileName.match(/\d{4}/);
+  return match ? match[0] : undefined;
+}
+
+async function loadDirectory(directory: string): Promise<LoadedDataset> {
+  const entries = await fs.promises.readdir(directory);
+  const imageFiles = entries.filter((entry) => /\.(png|jpe?g|bmp)$/i.test(entry)).sort();
+  if (imageFiles.length === 0) {
+    throw new Error(`No images found in ${directory}`);
+  }
+
+  const perDigit: DigitSample[] = [];
+  const files: LoadedFile[] = [];
+  for (const fileName of imageFiles) {
+    const filePath = path.join(directory, fileName);
+    const features = await extractDigits(filePath);
+    const expected = parseLabelFromFilename(fileName);
+    if (expected) {
+      features.forEach((feature, index) => {
+        perDigit.push({ label: Number(expected[index]!), feature });
+      });
+    }
+    files.push({ file: fileName, path: filePath, expected, features });
+  }
+
+  return { directory, perDigit, files };
+}
+
+function softmax(logits: Float64Array): Float64Array {
+  let max = -Infinity;
+  for (let i = 0; i < logits.length; i += 1) {
+    if (logits[i] > max) max = logits[i];
+  }
+
+  let sum = 0;
+  const output = new Float64Array(logits.length);
+  for (let i = 0; i < logits.length; i += 1) {
+    const exp = Math.exp(logits[i] - max);
+    output[i] = exp;
+    sum += exp;
+  }
+  for (let i = 0; i < output.length; i += 1) {
+    output[i] /= sum;
+  }
+  return output;
+}
+
+function shuffleInPlace<T>(array: T[]): void {
+  for (let i = array.length - 1; i > 0; i -= 1) {
+    const j = Math.floor(Math.random() * (i + 1));
+    [array[i], array[j]] = [array[j], array[i]];
+  }
+}
+
+function trainSoftmax(dataset: DigitSample[], inputSize: number): Float64Array {
+  const weights = new Float64Array(inputSize * CLASSES).fill(0);
+  for (let epoch = 0; epoch < MAX_EPOCHS; epoch += 1) {
+    shuffleInPlace(dataset);
+    let loss = 0;
+    for (const sample of dataset) {
+      const logits = new Float64Array(CLASSES);
+      for (let c = 0; c < CLASSES; c += 1) {
+        let sum = 0;
+        const offset = c * inputSize;
+        for (let i = 0; i < inputSize; i += 1) {
+          sum += weights[offset + i] * sample.feature[i];
+        }
+        logits[c] = sum;
+      }
+      const probs = softmax(logits);
+      loss += -Math.log(probs[sample.label] + 1e-9);
+
+      for (let c = 0; c < CLASSES; c += 1) {
+        const gradient = probs[c] - (c === sample.label ? 1 : 0);
+        const offset = c * inputSize;
+        for (let i = 0; i < inputSize; i += 1) {
+          const delta = gradient * sample.feature[i] + L2_LAMBDA * weights[offset + i];
+          weights[offset + i] -= LEARNING_RATE * delta;
+        }
+      }
+    }
+
+    if ((epoch + 1) % 100 === 0) {
+      const avgLoss = loss / dataset.length;
+      console.log(`epoch ${epoch + 1}: loss=${avgLoss.toFixed(6)}`);
+    }
+  }
+  return weights;
+}
+
+function predictDigit(weights: Float64Array, feature: Float64Array, inputSize: number): number {
+  let bestClass = 0;
+  let bestScore = -Infinity;
+  for (let c = 0; c < CLASSES; c += 1) {
+    let score = 0;
+    const offset = c * inputSize;
+    for (let i = 0; i < inputSize; i += 1) {
+      score += weights[offset + i] * feature[i];
+    }
+    if (score > bestScore) {
+      bestScore = score;
+      bestClass = c;
+    }
+  }
+  return bestClass;
+}
+
+async function evaluateDataset(
+  label: string,
+  files: LoadedFile[],
+  weights: Float64Array,
+  inputSize: number,
+  worker: Awaited<ReturnType<typeof createWorker>>,
+): Promise<RecognizedResult[]> {
+  let labeledCount = 0;
+  let correct = 0;
+  const results: RecognizedResult[] = [];
+
+  for (const file of files) {
+    const predictedDigits = file.features
+      .map((feature) => predictDigit(weights, feature, inputSize))
+      .join('');
+
+    let tesseractGuess: string | undefined;
+    try {
+      const { data } = await worker.recognize(file.path);
+      const cleaned = data.text.replace(/\D/g, '');
+      tesseractGuess = cleaned || undefined;
+    } catch (error) {
+      console.warn(`tesseract failed on ${file.file}:`, (error as Error).message);
+    }
+
+    if (file.expected) {
+      labeledCount += 1;
+      if (predictedDigits === file.expected) {
+        correct += 1;
+      }
+    }
+
+    results.push({
+      file: file.file,
+      expected: file.expected,
+      predicted: predictedDigits,
+      tesseract: tesseractGuess,
+    });
+  }
+
+  if (labeledCount > 0) {
+    console.log(`[${label}] 准确率：${correct}/${labeledCount}`);
+  } else {
+    console.log(`[${label}] 未提供标签，跳过准确率计算。`);
+  }
+
+  return results;
+}
+
+async function main() {
+  const args = process.argv.slice(2);
+
+  let trainDir = TRAIN_DIR_DEFAULT;
+  let validDir = VALID_DIR_DEFAULT;
+  let trainData: LoadedDataset | undefined;
+  let validData: LoadedDataset | undefined;
+
+  if (args.length > 0) {
+    const firstPath = path.resolve(args[0]);
+    if (!fs.existsSync(firstPath)) {
+      throw new Error(`指定的目录不存在：${firstPath}`);
+    }
+    const firstDataset = await loadDirectory(firstPath);
+    if (args.length > 1 || firstDataset.perDigit.length > 0) {
+      // 显式提供训练目录或该目录中包含标签，则视为训练集。
+      trainDir = firstPath;
+      trainData = firstDataset;
+      if (args.length > 1) {
+        validDir = path.resolve(args[1]);
+      }
+    } else {
+      // 仅提供一个无标签目录，视为验证集覆盖，同时保持默认训练集。
+      validDir = firstPath;
+      validData = firstDataset;
+    }
+  }
+
+  if (args.length > 1 && !fs.existsSync(validDir)) {
+    throw new Error(`验证目录不存在：${validDir}`);
+  }
+
+  if (!trainData) {
+    if (!fs.existsSync(trainDir)) {
+      throw new Error(`训练目录不存在：${trainDir}`);
+    }
+    trainData = await loadDirectory(trainDir);
+  }
+  if (trainData.perDigit.length === 0) {
+    throw new Error('训练集中未找到带标签的样本，无法训练模型。');
+  }
+
+  const inputSize = DIGIT_SIZE * DIGIT_SIZE;
+  const weights = trainSoftmax(trainData.perDigit, inputSize);
+
+  const worker = await createWorker('eng');
+  await worker.setParameters({
+    tessedit_char_whitelist: '0123456789',
+    tessedit_pageseg_mode: PSM.SINGLE_LINE,
+    user_defined_dpi: '300',
+  });
+
+  console.log('\n--- 训练集评估 ---');
+  const trainResults = await evaluateDataset('train', trainData.files, weights, inputSize, worker);
+
+  let validResults: RecognizedResult[] | undefined;
+  if (validDir && fs.existsSync(validDir)) {
+    try {
+      const validStats = validData ?? (await loadDirectory(validDir));
+      console.log('\n--- 验证集评估 ---');
+      validResults = await evaluateDataset('valid', validStats.files, weights, inputSize, worker);
+    } catch (error) {
+      if ((error as Error).message.includes('No images found')) {
+        console.log(`\n验证目录 ${validDir} 中未找到图片，跳过验证。`);
+      } else {
+        throw error;
+      }
+    }
+  } else if (validDir) {
+    console.log(`\n未找到验证目录 ${validDir}，仅输出训练集结果。`);
+  }
+
+  await worker.terminate();
+
+  const printResults = (title: string, results: RecognizedResult[]) => {
+    console.log(`\n${title}详细结果:`);
+    for (const result of results) {
+      const parts = [
+        `file=${result.file}`,
+        result.expected ? `expected=${result.expected}` : undefined,
+        `predicted=${result.predicted}`,
+        result.tesseract ? `tesseract=${result.tesseract}` : undefined,
+      ].filter(Boolean);
+      console.log(` - ${parts.join(' | ')}`);
+    }
+  };
+
+  printResults('训练集', trainResults);
+  if (validResults) {
+    printResults('验证集', validResults);
+  }
+}
+
+main().catch((error) => {
+  console.error(error);
+  process.exitCode = 1;
+});
diff --git a/todolist.md b/todolist.md
new file mode 100644
index 0000000..518200e
--- /dev/null
+++ b/todolist.md
@@ -0,0 +1,8 @@
+使用最合适的typescript开源 ocr 库，识别文件夹下的图片。
+
+图片上有两条弧形干扰线：一条白色干扰线在上，一条文字同色的干扰线在下。两条干扰线在图片中的位置基本不变。
+
+图片位 4 位阿拉伯数字。
+train文件: 文件名就是图片的四位数字，可作为训练校验用。
+valid文件: 文件名与图片内容无关，作为验证用。
+
diff --git a/train/0373.jpeg b/train/0373.jpeg
new file mode 100644
index 0000000..3fc514b
Binary files /dev/null and b/train/0373.jpeg differ
diff --git a/train/0462.jpeg b/train/0462.jpeg
new file mode 100644
index 0000000..55aa27f
Binary files /dev/null and b/train/0462.jpeg differ
diff --git a/train/0693.jpeg b/train/0693.jpeg
new file mode 100644
index 0000000..d18545a
Binary files /dev/null and b/train/0693.jpeg differ
diff --git a/train/1756.jpeg b/train/1756.jpeg
new file mode 100644
index 0000000..e2c6a80
Binary files /dev/null and b/train/1756.jpeg differ
diff --git a/train/2490.jpeg b/train/2490.jpeg
new file mode 100644
index 0000000..730dd6f
Binary files /dev/null and b/train/2490.jpeg differ
diff --git a/train/2797.jpeg b/train/2797.jpeg
new file mode 100644
index 0000000..b911535
Binary files /dev/null and b/train/2797.jpeg differ
diff --git a/train/4406.jpeg b/train/4406.jpeg
new file mode 100644
index 0000000..003bd5b
Binary files /dev/null and b/train/4406.jpeg differ
diff --git a/train/4459.jpeg b/train/4459.jpeg
new file mode 100644
index 0000000..28dd569
Binary files /dev/null and b/train/4459.jpeg differ
diff --git a/train/4705.jpeg b/train/4705.jpeg
new file mode 100644
index 0000000..4d76ec5
Binary files /dev/null and b/train/4705.jpeg differ
diff --git a/train/5009.jpeg b/train/5009.jpeg
new file mode 100644
index 0000000..6292f4a
Binary files /dev/null and b/train/5009.jpeg differ
diff --git a/train/5916.jpeg b/train/5916.jpeg
new file mode 100644
index 0000000..6eaaf30
Binary files /dev/null and b/train/5916.jpeg differ
diff --git a/train/5984.jpeg b/train/5984.jpeg
new file mode 100644
index 0000000..446f6ca
Binary files /dev/null and b/train/5984.jpeg differ
diff --git a/train/6117.jpeg b/train/6117.jpeg
new file mode 100644
index 0000000..53cbc17
Binary files /dev/null and b/train/6117.jpeg differ
diff --git a/train/6266.jpeg b/train/6266.jpeg
new file mode 100644
index 0000000..9ca30d7
Binary files /dev/null and b/train/6266.jpeg differ
diff --git a/train/6343.jpeg b/train/6343.jpeg
new file mode 100644
index 0000000..1e61a38
Binary files /dev/null and b/train/6343.jpeg differ
diff --git a/train/6820.jpeg b/train/6820.jpeg
new file mode 100644
index 0000000..34cdab4
Binary files /dev/null and b/train/6820.jpeg differ
diff --git a/train/6875.jpeg b/train/6875.jpeg
new file mode 100644
index 0000000..8b66abd
Binary files /dev/null and b/train/6875.jpeg differ
diff --git a/train/7151.jpeg b/train/7151.jpeg
new file mode 100644
index 0000000..bf71a43
Binary files /dev/null and b/train/7151.jpeg differ
diff --git a/train/7238.jpeg b/train/7238.jpeg
new file mode 100644
index 0000000..18fd70d
Binary files /dev/null and b/train/7238.jpeg differ
diff --git a/train/7733.jpeg b/train/7733.jpeg
new file mode 100644
index 0000000..d938b4e
Binary files /dev/null and b/train/7733.jpeg differ
diff --git a/train/8107.jpeg b/train/8107.jpeg
new file mode 100644
index 0000000..5878285
Binary files /dev/null and b/train/8107.jpeg differ
diff --git a/train/8324.jpeg b/train/8324.jpeg
new file mode 100644
index 0000000..a0262f4
Binary files /dev/null and b/train/8324.jpeg differ
diff --git a/train/8930.jpeg b/train/8930.jpeg
new file mode 100644
index 0000000..52a2353
Binary files /dev/null and b/train/8930.jpeg differ
diff --git a/train/9055.jpeg b/train/9055.jpeg
new file mode 100644
index 0000000..12a6948
Binary files /dev/null and b/train/9055.jpeg differ
diff --git a/train/9064.jpeg b/train/9064.jpeg
new file mode 100644
index 0000000..de6c995
Binary files /dev/null and b/train/9064.jpeg differ
diff --git a/train/9338.jpeg b/train/9338.jpeg
new file mode 100644
index 0000000..8f06747
Binary files /dev/null and b/train/9338.jpeg differ
diff --git a/train/9405.jpeg b/train/9405.jpeg
new file mode 100644
index 0000000..c036986
Binary files /dev/null and b/train/9405.jpeg differ
diff --git a/train/9685.jpeg b/train/9685.jpeg
new file mode 100644
index 0000000..2577928
Binary files /dev/null and b/train/9685.jpeg differ
diff --git a/train/9988.jpeg b/train/9988.jpeg
new file mode 100644
index 0000000..269c4ee
Binary files /dev/null and b/train/9988.jpeg differ
diff --git a/tsconfig.json b/tsconfig.json
new file mode 100644
index 0000000..8357513
--- /dev/null
+++ b/tsconfig.json
@@ -0,0 +1,14 @@
+{
+  "compilerOptions": {
+    "target": "es2019",
+    "module": "commonjs",
+    "moduleResolution": "node",
+    "types": ["node"],
+    "esModuleInterop": true,
+    "forceConsistentCasingInFileNames": true,
+    "skipLibCheck": true,
+    "strict": true,
+    "noEmit": true
+  },
+  "include": ["src"]
+}
diff --git a/valid/YZM-10.jpeg b/valid/YZM-10.jpeg
new file mode 100644
index 0000000..9fa83f3
Binary files /dev/null and b/valid/YZM-10.jpeg differ
diff --git a/valid/YZM-11.jpeg b/valid/YZM-11.jpeg
new file mode 100644
index 0000000..5e93340
Binary files /dev/null and b/valid/YZM-11.jpeg differ
diff --git a/valid/YZM-12.jpeg b/valid/YZM-12.jpeg
new file mode 100644
index 0000000..24e0f93
Binary files /dev/null and b/valid/YZM-12.jpeg differ
diff --git a/valid/YZM-13.jpeg b/valid/YZM-13.jpeg
new file mode 100644
index 0000000..3f1b8c5
Binary files /dev/null and b/valid/YZM-13.jpeg differ
diff --git a/valid/YZM-14.jpeg b/valid/YZM-14.jpeg
new file mode 100644
index 0000000..55e838b
Binary files /dev/null and b/valid/YZM-14.jpeg differ
diff --git a/valid/YZM-15.jpeg b/valid/YZM-15.jpeg
new file mode 100644
index 0000000..7c8d514
Binary files /dev/null and b/valid/YZM-15.jpeg differ
diff --git a/valid/YZM-16.jpeg b/valid/YZM-16.jpeg
new file mode 100644
index 0000000..ff3c9e9
Binary files /dev/null and b/valid/YZM-16.jpeg differ
diff --git a/valid/YZM-17.jpeg b/valid/YZM-17.jpeg
new file mode 100644
index 0000000..67b3e9c
Binary files /dev/null and b/valid/YZM-17.jpeg differ
diff --git a/valid/YZM-18.jpeg b/valid/YZM-18.jpeg
new file mode 100644
index 0000000..df75581
Binary files /dev/null and b/valid/YZM-18.jpeg differ
diff --git a/valid/YZM-19.jpeg b/valid/YZM-19.jpeg
new file mode 100644
index 0000000..6d11aee
Binary files /dev/null and b/valid/YZM-19.jpeg differ
diff --git a/valid/YZM-2.jpeg b/valid/YZM-2.jpeg
new file mode 100644
index 0000000..11c199f
Binary files /dev/null and b/valid/YZM-2.jpeg differ
diff --git a/valid/YZM-20.jpeg b/valid/YZM-20.jpeg
new file mode 100644
index 0000000..ea76fc3
Binary files /dev/null and b/valid/YZM-20.jpeg differ
diff --git a/valid/YZM-21.jpeg b/valid/YZM-21.jpeg
new file mode 100644
index 0000000..1250461
Binary files /dev/null and b/valid/YZM-21.jpeg differ
diff --git a/valid/YZM-22.jpeg b/valid/YZM-22.jpeg
new file mode 100644
index 0000000..1e06e4d
Binary files /dev/null and b/valid/YZM-22.jpeg differ
diff --git a/valid/YZM-23.jpeg b/valid/YZM-23.jpeg
new file mode 100644
index 0000000..10fb52d
Binary files /dev/null and b/valid/YZM-23.jpeg differ
diff --git a/valid/YZM-3.jpeg b/valid/YZM-3.jpeg
new file mode 100644
index 0000000..6dc57b0
Binary files /dev/null and b/valid/YZM-3.jpeg differ
diff --git a/valid/YZM-4.jpeg b/valid/YZM-4.jpeg
new file mode 100644
index 0000000..3c8164b
Binary files /dev/null and b/valid/YZM-4.jpeg differ
diff --git a/valid/YZM-5.jpeg b/valid/YZM-5.jpeg
new file mode 100644
index 0000000..934aba2
Binary files /dev/null and b/valid/YZM-5.jpeg differ
diff --git a/valid/YZM-6.jpeg b/valid/YZM-6.jpeg
new file mode 100644
index 0000000..5e0436f
Binary files /dev/null and b/valid/YZM-6.jpeg differ
diff --git a/valid/YZM-7.jpeg b/valid/YZM-7.jpeg
new file mode 100644
index 0000000..10b1a98
Binary files /dev/null and b/valid/YZM-7.jpeg differ
diff --git a/valid/YZM-8.jpeg b/valid/YZM-8.jpeg
new file mode 100644
index 0000000..2793ec0
Binary files /dev/null and b/valid/YZM-8.jpeg differ
diff --git a/valid/YZM-9.jpeg b/valid/YZM-9.jpeg
new file mode 100644
index 0000000..375d99f
Binary files /dev/null and b/valid/YZM-9.jpeg differ
diff --git a/valid/YZM.jpeg b/valid/YZM.jpeg
new file mode 100644
index 0000000..d416174
Binary files /dev/null and b/valid/YZM.jpeg differ