halcon深度学习异常检测

初学，记录学习过程
halcon深度学习例程dl_anomaly_detection_workflow，算法流程共分为5部分：
（0）准备阶段
准备输入、输出路径。
dev_update_off ()
dev_close_window ()
*设置随机种子
set_system (‘seed_rand’, 73)

HalconImages := ‘C:/Users/Administrator/Desktop’

• 指定图像类别所在的文件夹
  ImageDir := HalconImages+‘/leather’

• 限制读取的子文件夹
  ImageSubDirs := [‘good’,‘fold’,‘cut’,‘poke’,‘color’,‘glue’]
  *存放标签
  AnomalyLabelsDir := HalconImages+‘/labels/leather’
  OutputDir := ‘./anomaly_leather_data’
  *数据集预处理
  ExampleSpecificPreprocessing := true
  *图像缩放之后的大小，是32的倍数
  ImageWidth := 320
  ImageHeight := 320
  *复杂度，越大准确率越高，训练时间越长，代表模型处理更复杂缺陷的能力
  Complexity := 15
  （1）第一阶段：准备阶段
  GenParamDataset:=dict{image_sub_dirs:ImageSubDirs}
  *读取异常检测的深度学习数据集
  read_dl_dataset_anomaly (ImageDir, AnomalyLabelsDir, [], [], GenParamDataset, DLDataset)
  *拆分样本集为训练集（50%），验证集（20%），测试集（剩余的30%）
  split_dl_dataset (DLDataset, 50, 10, [])
  *读取MVTec提供的初始网络之一

• 加载预训练的模型、设置参数
  read_dl_model (‘initial_dl_anomaly_medium.hdl’, DLModelHandle)
  *设置之前定义的模型参数
  set_dl_model_param (DLModelHandle, ‘image_width’, ImageWidth)
  set_dl_model_param (DLModelHandle, ‘image_height’, ImageHeight)
  set_dl_model_param (DLModelHandle, ‘complexity’, Complexity)

• In this example, the anomaly detection model is trained on the CPU.
  query_available_dl_devices ([‘runtime’, ‘id’], [‘cpu’, 0], DLDevice)
  set_dl_model_param (DLModelHandle, ‘device’, DLDevice)
  *设置预处理参数，并进行预处理
  create_dl_preprocess_param_from_model (DLModelHandle, ‘constant_values’, ‘full_domain’, [], [], [], DLPreprocessParam)
  PreprocessSetting:=dict{overwrite_files:‘true’}
  preprocess_dl_dataset (DLDataset, OutputDir, DLPreprocessParam, PreprocessSetting, DLDatasetFileName)
  *获得样本集DLDataset中的样本
  if(ExampleSpecificPreprocessing)
  read_dl_samples (DLDataset, [0:|DLDataset.samples|-1], DLSampleBatch)
  *对图像背景进行处理
  preprocess_dl_samples_bottle(DLSampleBatch)
  write_dl_samples (DLDataset, [0:|DLDataset.samples|-1], DLSampleBatch, [], [])
  endif
  展示10个随机预处理后的DLSample
  WindowDict:=dict{}
  for Index := 0 to 9 by 1
  SampleIndex:=int(rand(1)|DLDataset.samples|)
  *读取深度学习样本
  read_dl_samples (DLDataset, SampleIndex, DLSample)
  *显示所有数据
  dev_display_dl_data (DLSample, [], DLSample, ‘anomaly_ground_truth’, [], WindowDict)

  dev_disp_text (‘Press Run (F5) to continue’, ‘window’, ‘bottom’, ‘right’, ‘black’, [], [])
  dev_set_window (WindowDict.anomaly_ground_truth[0])
  dev_disp_text (‘Preprocessed image’, ‘window’, ‘top’, ‘left’, ‘black’, [], [])
  stop()
  endfor

dev_close_window_dict (WindowDict)
stop()
（2）第二阶段：训练阶段
EnableDisplay:=true
*创建字典，并存储键-值对
TrainPramaAnomaly:=dict{}
TrainPramaAnomaly.error_threshold:=0.001
MaxNumEpochs:=30

TrainPramaAnomaly.domain_ratio:=0.25
TrainPramaAnomaly.regularization_noise:=0.01
*创建训练参数
create_dl_train_param (DLModelHandle, MaxNumEpochs, [], EnableDisplay, 73, ‘anomaly’, TrainPramaAnomaly, TrainParam)

train_dl_model (DLDataset, DLModelHandle, TrainParam, 0, TrainResults, TrainInfos, EvaluationInfos)
dev_disp_text (‘Press Run (F5) to continue’, ‘window’, ‘bottom’, ‘right’, ‘black’, [], [])
stop()
*
dev_close_window()

（3）第三阶段：评估阶段
*标准差因子（如果缺陷很小，推荐较大值）
StandardDeviationFactor:=1.0
set_dl_model_param (DLModelHandle, ‘standard_deviation_factor’, StandardDeviationFactor)

*计算阈值
GenParamThresholds:=dict{enable_display:true}
compute_dl_anomaly_thresholds (DLModelHandle, DLDataset, GenParamThresholds, AnomalySegmentationThreshold, AnomalyClassificationThresholds)
dev_disp_text (‘Press Run (F5) to continue’, ‘window’, ‘bottom’, ‘right’, ‘black’, [], [])
stop()

dev_close_window ()
*设置评估参数，在test集上进行评估
GenParamEvaluation:=dict{}
GenParamEvaluation.measures:=‘all’
GenParamEvaluation.anomaly_classification_thresholds:=AnomalyClassificationThresholds
evaluate_dl_model (DLDataset, DLModelHandle, ‘split’, ‘test’, GenParamEvaluation, EvaluationResult, EvalParams)
*要展示的参数
GenParamDisplay:=dict{}
*直方图、图例
GenParamDisplay.display_mode:=[‘score_histogram’,‘score_legend’]
WindowDict:=dict{}
*生成了一张直方图
dev_display_anomaly_detection_evaluation (EvaluationResult, EvalParams, GenParamDisplay, WindowDict)
dev_disp_text (‘Press Run (F5) to continue’, ‘window’, ‘bottom’, ‘right’, ‘black’, [], [])
stop()
*
dev_close_window_dict (WindowDict)

• 可视化precision，recall召回率，和confusion_matrix
  GenParamDisplay.display_mode:=[‘pie_charts_precision’,‘pie_charts_recall’,‘absolute_confusion_matrix’]
  ClassificationThresholdIndex:=|AnomalyClassificationThresholds|-1
  GenParamDisplay.classification_threshold_index:=ClassificationThresholdIndex
  WindowDict:=dict{}
  dev_display_anomaly_detection_evaluation (EvaluationResult, EvalParams, GenParamDisplay, WindowDict)
  dev_disp_text (‘Press Run (F5) to continue’, ‘window’, ‘bottom’, ‘right’, ‘black’, [], [])
  stop()

dev_close_window_dict (WindowDict)
get_dl_model_param (DLModelHandle, ‘meta_data’, MetaData)
MetaData.anomaly_segmentation_threshold:=AnomalySegmentationThreshold${}^{\prime }1.16e^{\prime }MetaData.anomal{y}_{c}lassificatio{n}_{t}hreshold:=AnomalyClassificationThresholds[ClassificationThresholdIndex]$‘1.16e’
set_dl_model_param (DLModelHandle, ‘meta_data’, MetaData)

write_dl_model (DLModelHandle, ‘model_final.hdl’)
（4）第四阶段：推理阶段
list_image_files (ImageDir+‘/’+ImageSubDirs, ‘default’, ‘recursive’, ImageFiles)
*打乱数据集
tuple_shuffle (ImageFiles, ImageFilesShuffled)
*设置阈值，模型训练后得到
get_dl_model_param (DLModelHandle, ‘meta_data’, MetaData)
InferenceClassificationThreshold:=number(MetaData.anomaly_classification_threshold)
InferenceSegmentationThreshold:=number(MetaData.anomaly_segmentation_threshold)

• 创建类别标签字典
  DLDatasetInfo := dict{class_names:[‘OK’,‘nok’],class_ids:[0,1]}
  WindowDict:=dict{}
  for IndexInference := 0 to min2(|ImageFilesShuffled|,20)-1 by 1
  read_image (Image, ImageFilesShuffled[IndexInference])
  gen_dl_samples_from_images (Image, DLSample)
  preprocess_dl_samples (DLSample, DLPreprocessParam)
  if(ExampleSpecificPreprocessing)
  preprocess_dl_samples_bottle (DLSample)
  endif
  apply_dl_model (DLModelHandle, DLSample, [], DLResult)
  threshold_dl_anomaly_results (InferenceSegmentationThreshold, InferenceClassificationThreshold, DLResult)

  dev_display_dl_data (DLSample, DLResult, DLDatasetInfo, [‘anomaly_result’,‘anomaly_image’], [], WindowDict)
  dev_disp_text (‘Press Run (F5) to continue’, ‘window’, ‘bottom’, ‘right’, ‘black’, [], [])
  stop()
  endfor
  dev_close_window_dict (WindowDict)