R-KDE#

Region Based Anomaly Detection With Real-Time Training and Analysis.

https://ieeexplore.ieee.org/abstract/document/8999287

class anomalib.models.image.rkde.lightning_model.Rkde(roi_stage=RoiStage.RCNN, roi_score_threshold=0.001, min_box_size=25, iou_threshold=0.3, max_detections_per_image=100, n_pca_components=16, feature_scaling_method=FeatureScalingMethod.SCALE, max_training_points=40000)#

Bases: MemoryBankMixin, AnomalyModule

Region Based Anomaly Detection With Real-Time Training and Analysis.

Parameters:

roi_stage (RoiStage, optional) – Processing stage from which rois are extracted. Defaults to RoiStage.RCNN.
roi_score_threshold (float, optional) – Mimumum confidence score for the region proposals. Defaults to 0.001.
min_size (int, optional) – Minimum size in pixels for the region proposals. Defaults to 25.
iou_threshold (float, optional) – Intersection-Over-Union threshold used during NMS. Defaults to 0.3.
max_detections_per_image (int, optional) – Maximum number of region proposals per image. Defaults to 100.
n_pca_components (int, optional) – Number of PCA components. Defaults to 16.
feature_scaling_method (FeatureScalingMethod, optional) – Scaling method applied to features before passing to KDE. Options are norm (normalize to unit vector length) and scale (scale to max length observed in training). Defaults to FeatureScalingMethod.SCALE.
max_training_points (int, optional) – Maximum number of training points to fit the KDE model. Defaults to 40000.

static configure_optimizers()#

RKDE doesn’t require optimization, therefore returns no optimizers.

Return type:: None

fit()#

Fit a KDE Model to the embedding collected from the training set.

Return type:: None

property learning_type: LearningType#

Return the learning type of the model.

Returns:: Learning type of the model.
Return type:: LearningType

property trainer_arguments: dict[str, Any]#

Return R-KDE trainer arguments.

Returns:: Arguments for the trainer.
Return type:: dict[str, Any]

training_step(batch, *args, **kwargs)#

Perform a training Step of RKDE. For each batch, features are extracted from the CNN.

Parameters:

batch (dict[str, str | torch.Tensor]) – Batch containing image filename, image, label and mask
args – Additional arguments.
kwargs – Additional keyword arguments.

Return type:

None

Returns:

Deep CNN features.

validation_step(batch, *args, **kwargs)#

Perform a validation Step of RKde.

Similar to the training step, features are extracted from the CNN for each batch.

Parameters:

batch (dict[str, str | torch.Tensor]) – Batch containing image filename, image, label and mask
args – Additional arguments.
kwargs – Additional keyword arguments.

Return type:

Union[Tensor, Mapping[str, Any], None]

Returns:

Dictionary containing probability, prediction and ground truth values.

Torch model for region-based anomaly detection.

class anomalib.models.image.rkde.torch_model.RkdeModel(roi_stage=RoiStage.RCNN, roi_score_threshold=0.001, min_box_size=25, iou_threshold=0.3, max_detections_per_image=100, n_pca_components=16, feature_scaling_method=FeatureScalingMethod.SCALE, max_training_points=40000)#

Bases: Module

Torch Model for the Region-based Anomaly Detection Model.

Parameters:

roi_stage (RoiStage, optional) – Processing stage from which rois are extracted. Defaults to RoiStage.RCNN.
roi_score_threshold (float, optional) – Mimumum confidence score for the region proposals. Defaults to 0.001.
min_size (int, optional) – Minimum size in pixels for the region proposals. Defaults to 25.
iou_threshold (float, optional) – Intersection-Over-Union threshold used during NMS. Defaults to 0.3.
max_detections_per_image (int, optional) – Maximum number of region proposals per image. Defaults to 100.
n_pca_components (int, optional) – Number of PCA components. Defaults to 16.
feature_scaling_method (FeatureScalingMethod, optional) – Scaling method applied to features before passing to KDE. Options are norm (normalize to unit vector length) and scale (scale to max length observed in training). Defaults to FeatureScalingMethod.SCALE.
max_training_points (int, optional) – Maximum number of training points to fit the KDE model. Defaults to 40000.

fit(embeddings)#

Fit the model using a set of collected embeddings.

Parameters:: embeddings (torch.Tensor) – Input embeddings to fit the model.
Return type:: bool
Returns:: Boolean confirming whether the training is successful.

forward(batch)#

Prediction by normality model.

Parameters:

batch (torch.Tensor) – Input images.

Returns:

The extracted features (when in training mode),: or the predicted rois and corresponding anomaly scores.

Return type:

Tensor | tuple[torch.Tensor, torch.Tensor]

Region-based Anomaly Detection with Real Time Training and Analysis.

Feature Extractor.

class anomalib.models.image.rkde.feature_extractor.FeatureExtractor#

Bases: Module

Feature Extractor module for Region-based anomaly detection.

forward(batch, rois)#

Perform a forward pass of the feature extractor.

Parameters:

batch (torch.Tensor) – Batch of input images of shape [B, C, H, W].
rois (torch.Tensor) – torch.Tensor of shape [N, 5] describing the regions-of-interest in the batch.

Returns:

torch.Tensor containing a 4096-dimensional feature vector for every RoI location.

Return type:

Tensor

Region-based Anomaly Detection with Real Time Training and Analysis.

Region Extractor.

class anomalib.models.image.rkde.region_extractor.RegionExtractor(stage=RoiStage.RCNN, score_threshold=0.001, min_size=25, iou_threshold=0.3, max_detections_per_image=100)#

Bases: Module

Extracts regions from the image.

Parameters:

stage (RoiStage, optional) – Processing stage from which rois are extracted. Defaults to RoiStage.RCNN.
score_threshold (float, optional) – Mimumum confidence score for the region proposals. Defaults to 0.001.
min_size (int, optional) – Minimum size in pixels for the region proposals. Defaults to 25.
iou_threshold (float, optional) – Intersection-Over-Union threshold used during NMS. Defaults to 0.3.
max_detections_per_image (int, optional) – Maximum number of region proposals per image. Defaults to 100.

forward(batch)#

Forward pass of the model.

Parameters:

batch (torch.Tensor) – Batch of input images of shape [B, C, H, W].

Raises:

ValueError – When stage is not one of rcnn or rpn.

Returns:

Predicted regions, tensor of shape [N, 5] where N is the number of predicted regions in the batch,: and where each row describes the index of the image in the batch and the 4 bounding box coordinates.

Return type:

Tensor

post_process_box_predictions(pred_boxes, pred_scores)#

Post-processes the box predictions.

The post-processing consists of removing small boxes, applying nms, and keeping only the k boxes with the highest confidence score.

Parameters:

pred_boxes (torch.Tensor) – Box predictions of shape (N, 4).
pred_scores (torch.Tensor) – torch.Tensor of shape () with a confidence score for each box prediction.

Returns:

Post-processed box predictions of shape (N, 4).

Return type:

list[torch.Tensor]

class anomalib.models.image.rkde.region_extractor.RoiStage(value, names=None, *, module=None, qualname=None, type=None, start=1, boundary=None)#

Bases: str, Enum

Processing stage from which rois are extracted.

R-KDE

Contents

R-KDE#