• 服务医学，基于目标检测模型实现细胞检测识别

    在我以往的认知里面，显微镜下面的世界是很神奇的存在，可能只平时接触到的绝大多数是都是宏观的世界吧，看到微观世界里面各色各样的生物、细胞就会觉得很神奇，电子显微镜往往都是医生来操作观察的，对于采样、病理切片等任务比较繁重的时候，人工就会显得捉襟见肘了，这时候AI就可以提供帮助了，基于大量的医学数据进行学习拟合，得到的模型甚至能够超越领域专家，这些在很多主营医疗业务的公司里面都得到的实现，本文今天并不是谈很宏大的东西，就是做细胞数据集的目标检测，医学相关的数据感觉很有意思，后面有机会我也会多去做这个领域的项目，

    首先看下效果图：

     这里面一共标注了三种细胞类型：

    1. 红细胞
    2. 血小板
    3. 白细胞

    英文表示分别为：

    1. RBC
    2. Platelets
    3. WBC

    接下来看下数据集：

     最近都在做一些轻量级的模型，这里模型选用的是MobileNetV2-YOLOv3-Lite的，详情如下所示：

    1. [net]
    2. # Training
    3. batch=128
    4. subdivisions=2
    5. width=320
    6. height=320
    7. channels=3
    8. momentum=0.9
    9. decay=4e-5
    10. angle=0
    11. saturation = 1.5
    12. exposure = 1.5
    13. hue=.1
    14.  
    15. learning_rate=0.001
    16. burn_in=1000
    17. max_batches = 50200
    18. policy=steps
    19. steps=40000,45000
    20. scales=.1,.1
    21.  
    22. [convolutional]
    23. filters=32
    24. size=3
    25. stride=2
    26. pad=1
    27. batch_normalize=1
    28. activation=relu
    29.  
    30. [convolutional]
    31. filters=32
    32. size=1
    33. stride=1
    34. pad=1
    35. batch_normalize=1
    36. activation=relu
    37.  
    38. [convolutional]
    39. filters=32
    40. size=3
    41. groups=32
    42. stride=1
    43. pad=1
    44. batch_normalize=1
    45. activation=relu
    46.  
    47. [convolutional]
    48. filters=16
    49. size=1
    50. stride=1
    51. pad=1
    52. batch_normalize=1
    53. activation=linear
    54.  
    55. [convolutional]
    56. filters=96
    57. size=1
    58. stride=1
    59. pad=1
    60. batch_normalize=1
    61. activation=relu
    62.  
    63. [convolutional]
    64. filters=96
    65. size=3
    66. groups=96
    67. stride=2
    68. pad=1
    69. batch_normalize=1
    70. activation=relu
    71.  
    72. [convolutional]
    73. filters=24
    74. size=1
    75. stride=1
    76. pad=1
    77. batch_normalize=1
    78. activation=linear
    79.  
    80. [convolutional]
    81. filters=144
    82. size=1
    83. stride=1
    84. pad=1
    85. batch_normalize=1
    86. activation=relu
    87.  
    88. [convolutional]
    89. filters=144
    90. size=3
    91. groups=144
    92. stride=1
    93. pad=1
    94. batch_normalize=1
    95. activation=relu
    96.  
    97. [convolutional]
    98. filters=24
    99. size=1
    100. stride=1
    101. pad=1
    102. batch_normalize=1
    103. activation=linear
    104.  
    105. [shortcut]
    106. from=-4
    107. activation=linear
    108.  
    109. [convolutional]
    110. filters=144
    111. size=1
    112. stride=1
    113. pad=1
    114. batch_normalize=1
    115. activation=relu
    116.  
    117. [convolutional]
    118. filters=144
    119. size=3
    120. groups=144
    121. stride=2
    122. pad=1
    123. batch_normalize=1
    124. activation=relu
    125.  
    126. [convolutional]
    127. filters=32
    128. size=1
    129. stride=1
    130. pad=1
    131. batch_normalize=1
    132. activation=linear
    133.  
    134. [convolutional]
    135. filters=192
    136. size=1
    137. stride=1
    138. pad=1
    139. batch_normalize=1
    140. activation=relu
    141.  
    142. [convolutional]
    143. filters=192
    144. size=3
    145. groups=192
    146. stride=1
    147. pad=1
    148. batch_normalize=1
    149. activation=relu
    150.  
    151. [convolutional]
    152. filters=32
    153. size=1
    154. stride=1
    155. pad=1
    156. batch_normalize=1
    157. activation=linear
    158.  
    159. [shortcut]
    160. from=-4
    161. activation=linear
    162.  
    163. [convolutional]
    164. filters=192
    165. size=1
    166. stride=1
    167. pad=1
    168. batch_normalize=1
    169. activation=relu
    170.  
    171. [convolutional]
    172. filters=192
    173. size=3
    174. groups=192
    175. stride=1
    176. pad=1
    177. batch_normalize=1
    178. activation=relu
    179.  
    180. [convolutional]
    181. filters=32
    182. size=1
    183. stride=1
    184. pad=1
    185. batch_normalize=1
    186. activation=linear
    187.  
    188. [shortcut]
    189. from=-4
    190. activation=linear
    191.  
    192. [convolutional]
    193. filters=192
    194. size=1
    195. stride=1
    196. pad=1
    197. batch_normalize=1
    198. activation=relu
    199.  
    200. [convolutional]
    201. filters=192
    202. size=3
    203. groups=192
    204. stride=1
    205. pad=1
    206. batch_normalize=1
    207. activation=relu
    208.  
    209. [convolutional]
    210. filters=64
    211. size=1
    212. stride=1
    213. pad=1
    214. batch_normalize=1
    215. activation=linear
    216.  
    217. [convolutional]
    218. filters=384
    219. size=1
    220. stride=1
    221. pad=1
    222. batch_normalize=1
    223. activation=relu
    224.  
    225. [convolutional]
    226. filters=384
    227. size=3
    228. groups=384
    229. stride=1
    230. pad=1
    231. batch_normalize=1
    232. activation=relu
    233.  
    234. [convolutional]
    235. filters=64
    236. size=1
    237. stride=1
    238. pad=1
    239. batch_normalize=1
    240. activation=linear
    241.  
    242. [shortcut]
    243. from=-4
    244. activation=linear
    245.  
    246. [convolutional]
    247. filters=384
    248. size=1
    249. stride=1
    250. pad=1
    251. batch_normalize=1
    252. activation=relu
    253.  
    254. [convolutional]
    255. filters=384
    256. size=3
    257. groups=384
    258. stride=1
    259. pad=1
    260. batch_normalize=1
    261. activation=relu
    262.  
    263. [convolutional]
    264. filters=64
    265. size=1
    266. stride=1
    267. pad=1
    268. batch_normalize=1
    269. activation=linear
    270.  
    271. [shortcut]
    272. from=-4
    273. activation=linear
    274.  
    275. [convolutional]
    276. filters=384
    277. size=1
    278. stride=1
    279. pad=1
    280. batch_normalize=1
    281. activation=relu
    282.  
    283. [convolutional]
    284. filters=384
    285. size=3
    286. groups=384
    287. stride=1
    288. pad=1
    289. batch_normalize=1
    290. activation=relu
    291.  
    292. [convolutional]
    293. filters=64
    294. size=1
    295. stride=1
    296. pad=1
    297. batch_normalize=1
    298. activation=linear
    299.  
    300. [shortcut]
    301. from=-4
    302. activation=linear
    303.  
    304. [convolutional]
    305. filters=384
    306. size=1
    307. stride=1
    308. pad=1
    309. batch_normalize=1
    310. activation=relu
    311.  
    312. [convolutional]
    313. filters=384
    314. size=3
    315. groups=384
    316. stride=2
    317. pad=1
    318. batch_normalize=1
    319. activation=relu
    320.  
    321. [convolutional]
    322. filters=96
    323. size=1
    324. stride=1
    325. pad=1
    326. batch_normalize=1
    327. activation=linear
    328.  
    329. [convolutional]
    330. filters=576
    331. size=1
    332. stride=1
    333. pad=1
    334. batch_normalize=1
    335. activation=relu
    336.  
    337. [convolutional]
    338. filters=576
    339. size=3
    340. groups=576
    341. stride=1
    342. pad=1
    343. batch_normalize=1
    344. activation=relu
    345.  
    346. [convolutional]
    347. filters=96
    348. size=1
    349. stride=1
    350. pad=1
    351. batch_normalize=1
    352. activation=linear
    353.  
    354. [shortcut]
    355. from=-4
    356. activation=linear
    357.  
    358. [convolutional]
    359. filters=576
    360. size=1
    361. stride=1
    362. pad=1
    363. batch_normalize=1
    364. activation=relu
    365.  
    366. [convolutional]
    367. filters=576
    368. size=3
    369. groups=576
    370. stride=1
    371. pad=1
    372. batch_normalize=1
    373. activation=relu
    374.  
    375. [convolutional]
    376. filters=96
    377. size=1
    378. stride=1
    379. pad=1
    380. batch_normalize=1
    381. activation=linear
    382.  
    383. [shortcut]
    384. from=-4
    385. activation=linear
    386.  
    387. [convolutional]
    388. filters=576
    389. size=1
    390. stride=1
    391. pad=1
    392. batch_normalize=1
    393. activation=relu
    394.  
    395. [convolutional]
    396. filters=576
    397. size=3
    398. groups=576
    399. stride=2
    400. pad=1
    401. batch_normalize=1
    402. activation=relu
    403.  
    404. [convolutional]
    405. filters=160
    406. size=1
    407. stride=1
    408. pad=1
    409. batch_normalize=1
    410. activation=linear
    411.  
    412. [convolutional]
    413. filters=960
    414. size=1
    415. stride=1
    416. pad=1
    417. batch_normalize=1
    418. activation=relu
    419.  
    420. [convolutional]
    421. filters=960
    422. size=3
    423. groups=960
    424. stride=1
    425. pad=1
    426. batch_normalize=1
    427. activation=relu
    428.  
    429. [convolutional]
    430. filters=160
    431. size=1
    432. stride=1
    433. pad=1
    434. batch_normalize=1
    435. activation=linear
    436.  
    437. [shortcut]
    438. from=-4
    439. activation=linear
    440.  
    441. [convolutional]
    442. filters=960
    443. size=1
    444. stride=1
    445. pad=1
    446. batch_normalize=1
    447. activation=relu
    448.  
    449. [convolutional]
    450. filters=960
    451. size=3
    452. groups=960
    453. stride=1
    454. pad=1
    455. batch_normalize=1
    456. activation=relu
    457.  
    458. [convolutional]
    459. filters=160
    460. size=1
    461. stride=1
    462. pad=1
    463. batch_normalize=1
    464. activation=linear
    465.  
    466. [shortcut]
    467. from=-4
    468. activation=linear
    469.  
    470. ### SPP ###
    471. [maxpool]
    472. stride=1
    473. size=3
    474.  
    475. [route]
    476. layers=-2
    477.  
    478. [maxpool]
    479. stride=1
    480. size=5
    481.  
    482. [route]
    483. layers=-4
    484.  
    485. [maxpool]
    486. stride=1
    487. size=9
    488.  
    489. [route]
    490. layers=-1,-3,-5,-6
    491.  
    492. ### End SPP ###
    493. #################################
    494. [convolutional]
    495. filters=288
    496. size=1
    497. stride=1
    498. pad=1
    499. batch_normalize=1
    500. activation=relu
    501.  
    502. [convolutional]
    503. filters=288
    504. size=3
    505. groups=288
    506. stride=1
    507. pad=1
    508. batch_normalize=1
    509. activation=relu
    510.  
    511. [convolutional]
    512. filters=96
    513. size=1
    514. stride=1
    515. pad=1
    516. batch_normalize=1
    517. activation=relu
    518.  
    519. [convolutional]
    520. filters=384
    521. size=1
    522. stride=1
    523. pad=1
    524. batch_normalize=1
    525. activation=relu
    526.  
    527. [convolutional]
    528. size=1
    529. stride=1
    530. pad=1
    531. filters=24
    532. activation=linear
    533.  
    534.  
    535. [yolo]
    536. mask = 3,4,5
    537. anchors = 26, 48,  67, 84,  72,175, 189,126, 137,236, 265,259
    538. classes=3
    539. num=6
    540. jitter=.1
    541. ignore_thresh = .5
    542. truth_thresh = 1
    543. random=1
    544. #################
    545. scale_x_y = 1.0
    546. iou_thresh=0.213
    547. cls_normalizer=1.0
    548. iou_normalizer=0.07
    549. iou_loss=ciou
    550. nms_kind=greedynms
    551. beta_nms=0.6
    552.  
    553. ##################################
    554. [route]
    555. layers= 65
    556.  
    557. [upsample]
    558. stride=2
    559.  
    560. [route]
    561. layers=-1,48
    562. #################################
    563. [convolutional]
    564. filters=80
    565. size=1
    566. stride=1
    567. pad=1
    568. batch_normalize=1
    569. activation=relu
    570.  
    571.  
    572. [convolutional]
    573. filters=288
    574. size=1
    575. stride=1
    576. pad=1
    577. batch_normalize=1
    578. activation=relu
    579.  
    580. [convolutional]
    581. filters=288
    582. size=3
    583. groups=288
    584. stride=1
    585. pad=1
    586. batch_normalize=1
    587. activation=relu
    588.  
    589. [convolutional]
    590. filters=192
    591. size=1
    592. stride=1
    593. pad=1
    594. batch_normalize=1
    595. activation=relu
    596.  
    597. [convolutional]
    598. filters=288
    599. size=1
    600. stride=1
    601. pad=1
    602. batch_normalize=1
    603. activation=relu
    604.  
    605. [convolutional]
    606. size=1
    607. stride=1
    608. pad=1
    609. filters=24
    610. activation=linear
    611.  
    612.  
    613. [yolo]
    614. mask = 0,1,2
    615. anchors = 26, 48,  67, 84,  72,175, 189,126, 137,236, 265,259
    616. classes=3
    617. num=6
    618. jitter=.1
    619. ignore_thresh = .5
    620. truth_thresh = 1
    621. random=1
    622. #################
    623. scale_x_y = 1.0
    624. iou_thresh=0.213
    625. cls_normalizer=1.0
    626. iou_normalizer=0.07
    627. iou_loss=ciou
    628. nms_kind=greedynms
    629. beta_nms=0.6

    同样一共需要修改两处参数配置

    【第一处】

     【第二处】

     

    主要就是修改里面classes和filters这两个关键字段即可：
          classes就是你要检测的目标对象数量，这里我要检测的目标对象数量是：3

          filters的数值计算公式为：  filters=(classes+5)*,在这里就是  （3+5）*3=24

    接下来编写cell.names和cell.data文件用于框架训练需要。

    cell.names如下所示：

    1. RBC
    2. Platelets
    3. WBC

    cell.data如下所示：

    1. classes= 3
    2. train  = /home/objDet/cell/train.txt
    3. valid  = /home/objDet/cell/test.txt
    4. names = /home/objDet/cell/x.names
    5. backup = /home/objDet/cell/model/yolo

    基于Darknet框架可以一键启动训练，命令如下：

    1. chmod +x darknet
    2. nohup ./darknet detector train x.data MobileNetV2-YOLOv3-Lite.cfg >> yolo.out &

    入股对数据集格式有疑问的可以翻看前面的文章，前面已经详细介绍过了，所以后面的文章就不在赘述了。

    随机选择数据测试结果如下：

     这里训练完成后， 我同样对整个训练过程的日志进行了可视化

    我在上一篇文章《AI助力智能安检，基于目标检测模型实现X光安检图像智能检测分析》中，编写了原始日志的解析代码，如下：

    1. def parseLogs(data="yolo.out"):
    2.     """
    3.     解析原始日志
    4.     """
    5.     with open(data) as f:
    6.         data_list=[one.strip() for one in f.readlines() if one.strip()]
    7.     print("data_list_length: ", len(data_list))
    8.     start=0
    9.     for i in range(len(data_list)):
    10.         if "hours left" in data_list[i]:
    11.             start=i
    12.             break
    13.     print("start: ", start)
    14.     s=start+1
    15.     res_list=[]
    16.     while s<len(data_list):
    17.         if "hours left" in data_list[s]:
    18.             one_epoch=data_list[start:s]
    19.             res_list.append(one_epoch)
    20.             start=s
    21.             s+=1
    22.         else:
    23.             s+=1
    24.     print("res_list_length: ", len(res_list))
    25.     for one_line in res_list[0]:
    26.         print(one_line)

    今天发现了可以有更简单的写法，代码如下：

    1. def extractNeedLog(log_file,new_log_file,key_word):
    2.     """
    3.     提取所需要的日志数据
    4.     """
    5.     with open(log_file, 'r') as f:
    6.         with open(new_log_file, 'w') as train_log:
    7.             for line in f:
    8.                 if key_word in line:
    9.                     train_log.write(line)
    10.     f.close()
    11.     train_log.close()

    因为Darknet框架的训练日志是非常规则的文本形式，可以根据简单的关键词匹配即可完成所需信息的提取操作，不需要像我之前那样先分离了单个epoch的日志，之后再二次提取。

    分离提取了loss日志和metric日志，这里对其进行可视化如下：

    loss曲线

     学习率曲线

     时耗曲线

     图像累积量曲线

     其实这个没有必要可视化，简单分析就知道是一条直线了，因为每次选取的图片数是固定的，乘以总次数自然就是一条正比例直线了。

    剩余时间曲线

     这个是模型训练过程中不断打印输出的剩余训练时间的曲线，总体趋势降低就是对的了，中间有起伏跟服务器硬件有关系。

    接下来是训练过程中的一些评估指标的可视化。

     我是把觉得能可视化的都给可视化了，其实真实可能并不需要这么多，只需要关注几个重点需要关注的指标即可，比如：IOU，Obj、Class等等。

  • 相关阅读:
    Windows系统补丁管理工具
    现网常见问题及解决思路
    怎么把一个excel拆分成多个
    如何办理软件企业认定证书
    MFC Windows 程序设计[310]之混搭个性按钮组群(附源码)
    LeetCode第 85 场双周赛总结
    汇编语言学习笔记及总结
    【Leetcode】1381. Design a Stack With Increment Operation
    mysql是怎么运行的-笔记
    【Java】应用层协议HTTP和HTTPS
  • 原文地址：https://blog.csdn.net/Together_CZ/article/details/126041456