Эхлэл Бүгдийг харах Тусламж
Нэвтрэх
zijun.wang
/
PokemonCardSearch
1
0
Салаа 0
Файлууд Асуудлууд 0 Хуулах хүсэлтүүд 0 Мэдлэгийн сан
Салаа: master
Салаанууд Тагууд
PokemonCardSear...
/
code
/
pokemon
/
pokemon_report.py

pokemon_report.py 14 KB
Байнгын холболт Түүх Анхны өгөгдөл

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242 
  1. # -*- coding: utf-8 -*-
    2. 

  2. """
    3. 

  3. pokemon_report.py - Pokemon 评级卡识别 技术汇报 HTML (学习 report.html 风格)
    4. 

  4. 

  5. 主题: 评级卡区域识别 (best.pt 检测评级公司 + PaddleOCR 识别分数/编号)
    6. 

  6. 不含卡牌匹配(DINOv2)展示, 仅评级。
    7. 

  7. 运行: pytorch/任意有 cv2 环境, 在服务器跑(图在服务器)
    8. 

  8. """
    9. 

  9. import os, sys, csv, json, hashlib, random, base64
    10. 

  10. from collections import Counter
    11. 

  11. sys.stdout.reconfigure(encoding="utf-8")
    12. 

  12. 

  13. OUT_DIR = "/home/martin/顾工交接/wzj/images_pokemon_out"
    14. 

  14. random.seed(42)
    15. 

  15. 

  16. 

  17. def b64(img, h=240, q=78):
    18. 

  18.     import cv2
    19. 

  19.     if img is None:
    20. 

  20.         return ""
    21. 

  21.     if img.shape[0] != h:
    22. 

  22.         sc = h / img.shape[0]
    23. 

  23.         img = cv2.resize(img, (max(1, int(img.shape[1] * sc)), h))
    24. 

  24.     ok, buf = cv2.imencode(".jpg", img, [cv2.IMWRITE_JPEG_QUALITY, q])
    25. 

  25.     return "data:image/jpeg;base64," + base64.b64encode(buf.tobytes()).decode() if ok else ""
    26. 

  26. 

  27. 

  28. def load(p):
    29. 

  29.     import cv2
    30. 

  30.     return cv2.imread(p) if p and os.path.exists(p) else None
    31. 

  31. 

  32. 

  33. print("[1] 读数据 ...", flush=True)
    34. 

  34. rows = list(csv.DictReader(open(f"{OUT_DIR}/match_result_sorted.csv", encoding="utf-8-sig")))
    35. 

  35. ocr = json.load(open(f"{OUT_DIR}/ocr_result.json", encoding="utf-8"))
    36. 

  36. print(f"    {len(rows)} 张", flush=True)
    37. 

  37. 

  38. # ---- 评级统计 ----
    39. 

  39. graded = [r for r in rows if r["评级公司"] != "非评级卡"]
    40. 

  40. ungraded = [r for r in rows if r["评级公司"] == "非评级卡"]
    41. 

  41. correct = [r for r in graded if r["评级公司"] == r["true_label"]]
    42. 

  42. wrong = [r for r in graded if r["评级公司"] != r["true_label"]]
    43. 

  43. acc = len(correct) / len(graded) * 100
    44. 

  44. 

  45. COMPANIES = ["PSA", "BGS", "CGC", "SGC"]
    46. 

  46. comp_stat = {}
    47. 

  47. for cp in COMPANIES:
    48. 

  48.     sub = [r for r in rows if r["true_label"] == cp]
    49. 

  49.     c = sum(1 for r in sub if r["评级公司"] == cp)
    50. 

  50.     comp_stat[cp] = (c, len(sub), len(sub) - c)  # 正确, 总数, 误识别/漏检
    51. 

  51. 

  52. # 误识别混淆: 真实 -> 识别成什么
    53. 

  53. confuse = Counter((r["true_label"], r["评级公司"]) for r in wrong)
    54. 

  54. 

  55. # ---- OCR 分数/编号统计 ----
    56. 

  56. score_cnt = Counter(r["评级分数"] for r in graded)
    57. 

  57. n_serial = sum(1 for r in graded if r["serial_no"])
    58. 

  58. n_authentic = sum(1 for r in graded if r["评级分数"] == "authentic")
    59. 

  59. n_score_num = sum(1 for r in graded if r["评级分数"] and r["评级分数"] != "authentic")
    60. 

  60. 

  61. print(f"[2] 评级准确率 {acc:.2f}% (评级卡{len(graded)}, 非评级卡{len(ungraded)}, 误识别{len(wrong)})", flush=True)
    62. 

  62. print(f"    分数识别: 数字{n_score_num} authentic{n_authentic} | serial {n_serial}/{len(graded)}", flush=True)
    63. 

  63. 

  64. 

  65. # ---- 案例渲染 ----
    66. 

  66. def render(r, tag):
    67. 

  67.     fld, name = r["true_label"], r["filename"]
    68. 

  68.     lines = ocr.get(name, {}).get("lines", [])
    69. 

  69.     p_anno = b64(load(f"{OUT_DIR}/predict/{fld}/{name}"), h=280, q=82)
    70. 

  70.     p_crop = b64(load(f"{OUT_DIR}/crops/{fld}_{name}"), h=120, q=85)
    71. 

  71.     ocr_html = "<br>".join(lines[:14]) if lines else '<i style="color:#aaa">(OCR 无文本)</i>'
    72. 

  72.     ok_mark = "✅" if r["评级公司"] == r["true_label"] else "❌"
    73. 

  73.     return (f'<div class="case"><div class="ttl">{tag} <b>{name}</b> '
    74. 

  74.             f'<span class="meta">真实=<b>{fld}</b> → 识别=<b style="color:#3b82f6">{r["评级公司"]}</b> {ok_mark} | '
    75. 

  75.             f'分数=<b>{r["评级分数"]}</b> | 序号={r["serial_no"] or "-"}</span></div>'
    76. 

  76.             f'<div class="row"><div class="q"><img src="{p_anno}"><small>best.pt 评级检测</small></div>'
    77. 

  77.             f'<div class="q"><img src="{p_crop}"><small>评级标签条(OCR 输入)</small></div>'
    78. 

  78.             f'<div class="ocr"><div class="ocr-t">OCR 分行文本</div><pre>{ocr_html}</pre></div></div></div>')
    79. 

  79. 

  80. 

  81. def sample(lst, n):
    82. 

  82.     return random.sample(lst, min(n, len(lst))) if lst else []
    83. 

  83. 

  84. 

  85. cases = [("✅ 识别正确（评级公司=真实类别）", "#2e7d32", sample(correct, 60)),
    86. 

  86.          ("❌ 误识别（评级公司≠真实类别）", "#d32f2f", wrong),  # 全部(数量少)
    87. 

  87.          ("○ 非评级卡（best.pt 未检出评级框）", "#9e9e9e", ungraded)]
    88. 

  88. 

  89. # ---- HTML ----
    90. 

  90. print("[3] 生成 HTML ...", flush=True)
    91. 

  91. H = ['<!DOCTYPE html><html><head><meta charset="utf-8"><title>Pokemon 评级卡识别 技术汇报</title><style>']
    92. 

  92. H.append(
    93. 

  93.     'body{font-family:-apple-system,"Microsoft YaHei",sans-serif;margin:0;background:#f4f5f7;color:#222;}'
    94. 

  94.     'h1{background:linear-gradient(90deg,#1565c0,#7e57c2);color:#fff;padding:22px 24px;margin:0;font-size:22px;}'
    95. 

  95.     '.sub{color:#fff;background:#37474a;padding:8px 24px;font-size:13px;}'
    96. 

  96.     'h2{margin:24px 16px 10px;border-left:5px solid #1565c0;padding-left:10px;}'
    97. 

  97.     '.wrap{max-width:1300px;margin:0 auto;padding:0 16px 40px;}'
    98. 

  98.     '.kpis{display:flex;gap:12px;flex-wrap:wrap;margin:16px 0;}'
    99. 

  99.     '.kpi{flex:1;min-width:150px;background:#fff;border-radius:8px;padding:14px 18px;box-shadow:0 1px 3px rgba(0,0,0,.1);}'
    100. 

  100.     '.kpi .v{font-size:26px;font-weight:bold;} .kpi .l{color:#666;font-size:12px;margin-top:2px;}'
    101. 

  101.     '.panel{background:#fff;border-radius:8px;padding:14px 20px;margin:10px 0;box-shadow:0 1px 3px rgba(0,0,0,.08);}'
    102. 

  102.     'table{border-collapse:collapse;width:100%;font-size:14px;} th,td{border:1px solid #eee;padding:7px 10px;text-align:left;}'
    103. 

  103.     'th{background:#eef3ff;} .bar{height:16px;background:#1565c0;border-radius:3px;display:inline-block;vertical-align:middle;}'
    104. 

  104.     '.pipeline{display:flex;gap:8px;flex-wrap:wrap;} .step{background:#eef;border:1px solid #cdd;border-radius:6px;padding:8px 12px;font-size:13px;}'
    105. 

  105.     '.case{background:#fff;margin:10px 0;padding:12px;border-radius:7px;box-shadow:0 1px 3px rgba(0,0,0,.12);} '
    106. 

  106.     '.ttl{font-size:13px;margin-bottom:8px;} .meta{color:#666;font-size:12px;}'
    107. 

  107.     '.row{display:flex;gap:10px;align-items:flex-start;}'
    108. 

  108.     '.q{border:1px solid #ddd;padding:2px;border-radius:4px;text-align:center;background:#fafafa;} '
    109. 

  109.     '.q img{height:280px;display:block;} .q small{color:#888;font-size:11px;}'
    110. 

  110.     '.ocr{flex:1;background:#263238;color:#eee;border-radius:5px;padding:8px 12px;font-size:11px;min-width:200px;} '
    111. 

  111.     '.ocr-t{color:#80cbc4;font-size:11px;margin-bottom:4px;} .ocr pre{margin:0;white-space:pre-wrap;font-family:Consolas,monospace;line-height:1.5;}'
    112. 

  112.     '</style></head><body>')
    113. 

  113. 

  114. H.append('<h1> Pokemon 评级卡识别 · 技术汇报</h1>'
    115. 

  115.          '<div class="sub">主题：评级公司识别（PSA/BGS/CGC/SGC）+ 评级分数与编号 OCR ｜ '
    116. 

  116.          '数据集 images_pokemon 2000 张 ｜ 2026-07-03</div><div class="wrap">')
    117. 

  117. 

  118. # KPI
    119. 

  119. H.append('<div class="kpis">'
    120. 

  120.          f'<div class="kpi"><div class="v">{len(rows)}</div><div class="l">测试图片总数</div></div>'
    121. 

  121.          f'<div class="kpi"><div class="v" style="color:#2e7d32">{acc:.2f}%</div><div class="l">评级公司识别准确率<br>(仅评级卡 {len(graded)} 张)</div></div>'
    122. 

  122.          f'<div class="kpi"><div class="v" style="color:#d32f2f">{len(wrong)}</div><div class="l">误识别数量</div></div>'
    123. 

  123.          f'<div class="kpi"><div class="v" style="color:#9e9e9e">{len(ungraded)}</div><div class="l">非评级卡(未检出)</div></div>'
    124. 

  124.          f'<div class="kpi"><div class="v">{n_score_num}</div><div class="l">OCR 识别出数字分数</div></div>'
    125. 

  125.          f'<div class="kpi"><div class="v">{n_serial}/{len(graded)}</div><div class="l">编号 serial_no 识别率</div></div>'
    126. 

  126.          '</div>')
    127. 

  127. 

  128. # Pipeline
    129. 

  129. H.append('<h2>🔗 评级识别流程</h2><div class="panel"><div class="pipeline">'
    130. 

  130.          '<span class="step">① <b>best.pt (YOLO26)</b><br>在原图检测评级标签条<br>→ 输出评级公司 (PSA/BGS/CGC/SGC)</span>'
    131. 

  131.          '<span class="step">② 裁出评级标签条区域<br>(best.pt 检测框)</span>'
    132. 

  132.          '<span class="step">③ <b>PaddleOCR</b><br>识别标签条分行文本<br>→ 评级分数 + serial_no</span>'
    133. 

  133.          '<span class="step">④ 规则解析<br>分数限定 1-10 / 编号≥6位数字<br>评级卡无分数→authentic</span>'
    134. 

  134.          '</div></div>')
    135. 

  135. 

  136. # 评级公司准确率
    137. 

  137. H.append('<h2>📊 各评级公司识别准确率</h2><div class="panel"><table>'
    138. 

  138.          '<tr><th>真实类别</th><th>样本数</th><th>正确识别</th><th>误识别/漏检</th><th>准确率</th><th>可视化</th></tr>')
    139. 

  139. for cp in COMPANIES:
    140. 

  140.     c, t, w = comp_stat[cp]
    141. 

  141.     pct = c / t * 100 if t else 0
    142. 

  142.     H.append(f'<tr><td><b>{cp}</b></td><td>{t}</td><td>{c}</td><td>{w}</td>'
    143. 

  143.              f'<td><b>{pct:.2f}%</b></td><td><span class="bar" style="width:{pct*0.9}px"></span></td></tr>')
    144. 

  144. H.append(f'<tr style="background:#eef3ff"><td><b>合计(仅评级卡)</b></td><td>{len(graded)}</td>'
    145. 

  145.          f'<td>{len(correct)}</td><td>{len(wrong)}</td><td><b>{acc:.2f}%</b></td><td></td></tr></table></div>')
    146. 

  146. 

  147. # 误识别混淆
    148. 

  148. if wrong:
    149. 

  149.     H.append('<h2>❌ 误识别混淆（真实 → 误识别为）</h2><div class="panel"><table>'
    150. 

  150.              '<tr><th>真实类别</th><th>误识别为</th><th>数量</th></tr>')
    151. 

  151.     for (t, p), n in confuse.most_common():
    152. 

  152.         H.append(f'<tr><td>{t}</td><td><b style="color:#d32f2f">{p or "(空)"}</b></td><td>{n}</td></tr>')
    153. 

  153.     H.append(f'</table><p style="color:#666;font-size:13px;margin-top:8px">'
    154. 

  154.              f'共 {len(wrong)} 张误识别（评级卡中）。另 {len(ungraded)} 张非评级卡未检出评级框（不计入准确率）。</p></div>')
    155. 

  155. 

  156. # OCR 分数分析
    157. 

  157. H.append('<h2>🔢 OCR 评级分数与编号识别</h2><div class="panel">'
    158. 

  158.          f'<p>评级卡 {len(graded)} 张中：识别出数字分数 <b>{n_score_num}</b> 张，'
    159. 

  159.          f'评级卡但分数 OCR 漏识(填 authentic) <b>{n_authentic}</b> 张，'
    160. 

  160.          f'编号 serial_no(≥6位) 识别 <b>{n_serial}/{len(graded)} = {n_serial/len(graded)*100:.1f}%</b>。</p>'
    161. 

  161.          '<table><tr><th>评级分数</th><th>数量</th><th>占比</th></tr>')
    162. 

  162. for s, n in score_cnt.most_common(15):
    163. 

  163.     H.append(f'<tr><td><b>{s or "空"}</b></td><td>{n}</td><td>{n/len(graded)*100:.1f}%</td></tr>')
    164. 

  164. H.append('</table></div>')
    165. 

  165. 

  166. # 已知局限模块
    167. 

  167. H.append('''
    168. 

  168. <h2 style="border-color:#ef6c00;color:#ef6c00">⚠️ 已知局限：OCR 评级分数解析有待提高</h2>
    169. 

  169. <div class="panel" style="border-left:4px solid #ef6c00">
    170. 

  170. <h3 style="margin-top:0">当前判断逻辑（启发式，较粗糙）</h3>
    171. 

  171. <pre style="background:#263238;color:#eee;padding:10px;border-radius:5px;font-size:12px">for ln in lines:                     # 从上到下逐行扫
    172. 

  172.     if 纯数字 and 1.0 ≤ 该数 ≤ 10.0:
    173. 

  173.         return ln                    # 返回第一个命中的</pre>
    174. 

  174. <p>即"取 OCR 文本里<b>第一个</b>落在 1-10 范围的纯数字行"，完全靠行序，
    175. 

  175. <b>不理解标签布局、不认 BGS 四维分数、不区分评级公司、不看语义</b>。</p>
    176. 

  176. 

  177. <h3>主要问题</h3>
    178. 

  178. <ul>
    179. 

  179. <li><b>BGS 四维分数混淆</b>：BGS 标签有 CENTERING/CORNERS/EDGES/SURFACE 四个维度分 + 一个总分，共 5 个 1-10 数字。
    180. 

  180. 当前逻辑会把"第一个碰到的"当总分——可能撞上某个维度分，而非真实总分。</li>
    181. 

  181. <li><b>不区分评级公司</b>：PSA/CGC/SGC 是单分数标签（相对好办），BGS 是多分数（容易混），代码对四家一视同仁。</li>
    182. 

  182. <li><b>依赖 OCR 行序</b>：PaddleOCR 行序按文字框位置排，标签稍转一下、OCR 漏一行，行序就变，结果就漂。</li>
    183. 

  183. <li><b>卡牌信息数字干扰</b>：年份(2025)、编号(#020)、发行量(3/10) 等若被 OCR 拆成纯数字行，可能被误判为分数。</li>
    184. 

  184. </ul>
    185. 

  185. 

  186. <h3>具体案例：为何这张 BGS 判 10 而非 9.5</h3>
    187. 

  187. <div class="row" style="margin:8px 0">
    188. 

  188. <div class="ocr" style="min-width:320px">
    189. 

  189. <div class="ocr-t">OCR 分行文本（节选）</div>
    190. 

  190. <pre>2025 POKEMON MEGA PROMOS
    191. 

  191. <span style="color:#ffd54f">10            ← 第2行, 第一个1-10纯数字 → 判为分数</span>
    192. 

  192. B
    193. 

  193. #020MP PIKACHU
    194. 

  194. MCDONALD'S JAPAN
    195. 

  195. PURCHASE INCENTIVE
    196. 

  196. CENTERING
    197. 

  197. <span style="color:#80cbc4">9.5          ← CENTERING 维度分, 但已在10之后, 轮不到</span>
    198. 

  198. CORNERS 10
    199. 

  199. PRISTINE
    200. 

  200. BECKETT
    201. 

  201. EDGES 10
    202. 

  202. SURFACE 10</pre>
    203. 

  203. </div>
    204. 

  204. <div style="flex:1;font-size:13px">
    205. 

  205. <p><b>10</b> 在第 2 行、<b>9.5</b> 在第 9 行，逻辑先碰到 10 就返回——
    206. 

  206. 纯粹"谁排前面"，不是因为 10 更像总分。</p>
    207. 

  207. <p>旁证矛盾：文本出现 <code>PRISTINE</code>（BGS 黑标，对应总分 10），
    208. 

  208. 但 <code>CENTERING 9.5</code>（PRISTINE 要求四维全 10）→
    209. 

  209. 要么总分其实是 9.5（OCR 误读描述），要么行序错位。<b>无 ground truth 无法定论</b>。</p>
    210. 

  210. </div>
    211. 

  211. </div>
    212. 

  212. 

  213. <h3>改进方向</h3>
    214. 

  214. <ol>
    215. 

  215. <li><b>按公司分支策略</b>：BGS 专门定位"四维分项之外的总分"（位置先验 + 排除 CENTERING/CORNERS/EDGES/SURFACE 紧邻数字）；
    216. 

  216. PSA/CGC/SGC 取唯一主分数。</li>
    217. 

  217. <li><b>分数描述交叉校验</b>：PRISTINE→10 / GEM MT→10 / MINT→9 / NM-MT→8.5；分数与描述不符时报警或以描述反推。</li>
    218. 

  218. <li><b>布局定位</b>：best.pt 已框出评级标签条，在框内按区域定位总分（BGS 总分在固定格子），而非全文扫描。</li>
    219. 

  219. <li><b>专项标注+训练</b>：标几百张评级分数 ground truth，训小模型/规则，最稳但有标注成本。</li>
    220. 

  220. </ol>
    221. 

  221. </div>
    222. 

  222. ''')
    223. 

  223. 

  224. # 案例
    225. 

  225. H.append('<h2>🖼 识别案例可视化</h2>'
    226. 

  226.          '<p style="color:#666;font-size:13px;margin:0 16px;">每张：best.pt 评级检测图 ｜ 评级标签条(OCR 输入) ｜ OCR 分行文本 + 识别结果</p>')
    227. 

  227. for title, color, items in cases:
    228. 

  228.     if not items:
    229. 

  229.         continue
    230. 

  230.     total = len(correct if "正确" in title else wrong if "误" in title else ungraded)
    231. 

  231.     H.append(f'<h2 style="border-color:{color};color:{color}">{title} <small>(展示 {len(items)} / 共 {total} 张)</small></h2>')
    232. 

  232.     for r in items:
    233. 

  233.         try:
    234. 

  234.             H.append(render(r, f'<span class="tag" style="color:{color}">●</span>'))
    235. 

  235.         except Exception as e:
    236. 

  236.             H.append(f'<div class="case">[{r["filename"]} 渲染失败: {str(e)[:50]}]</div>')
    237. 

  237. 

  238. H.append('</div></body></html>')
    239. 

  239. out = f"{OUT_DIR}/pokemon_report.html"
    240. 

  240. with open(out, "w", encoding="utf-8") as f:
    241. 

  241.     f.write("".join(H))
    242. 

  242. print(f"[done] -> {out} ({os.path.getsize(out)/1e6:.1f} MB)", flush=True)
    243.