east_ocr

Custom node for OCR implementation with east-resnet50 and crnn models

This custom node analyses the response of east-resnet50 model. Based on the inference results and the original image, it generates a list of detected boxes for text recognition. Each image in the output will be resized to the predefined target size to fit the next inference model in the DAG pipeline. Additionally to the detected text boxes, in the two additional outputs are returned their coordinates with information about geometry and confidence levels for the filtered list of detections.

Building custom node library

You can build the shared library of the custom node simply by running command in this custom node folder context:

make

It will compile the library inside a docker container and save the results in lib folder.

Custom node inputs

Input name	Description	Shape	Precision
image	Input image in an array format. Only batch size 1 is supported and images must have 3 channels. Resolution is configurable via parameters original_image_width and original_image_height.	1,3,H,W	FP32
scores	east-resnet50 model output feature_fusion/Conv_7/Sigmoid	1,1,256,480	FP32
geometry	east-resnet50 model output feature_fusion/concat_3	1,5,256,480	FP32

Custom node outputs

Output name	Description	Shape	Precision
text_images	Returns images representing detected text boxes. Boxes are filtered based on confidence_threshold and overlap_threshold params. Resolution is defined by the node parameters. All images are in a single batch. Batch size depend on the number of detected objects.	N,1,C,H,W	FP32
text_coordinates	For every detected box N the following info is added: x coordinate for the box center, y coordinate for the box center, box original width, box original height	N,1,4	I32
confidence_levels	For every detected box N information about score result	N,1,1	FP32

Custom node parameters

Parameter	Description	Default	Required
original_image_width	Required input image width		✓
original_image_height	Required input image height		✓
target_image_width	Target width of the text boxes in output. Boxes in the original image will be resized to that value.		✓
target_image_height	Target width of the text boxes in output. Boxes in the original image will be resized to that value.		✓
convert_to_gray_scale	Defines if output images should be in grayscale or in color	false
confidence_threshold	Number in a range of 0-1		✓
overlap_threshold	a ratio in a range of 0-1 for non-max suppression algorithm. Defines the overlapping ratio to reject detection as duplicated	0.3
debug	Defines if debug messages should be displayed	false
max_output_batch	Prevents too big batches with incorrect confidence level. It can avoid exceeding RAM resources	100
box_width_adjustment	Horizontal size expansion level for text images to compensate cut letter. Letters might be cut on the edges in case of the EAST model accuracy problems. That parameter defines how much horizontal size should be expanded comparing to the original width	0
box_height_adjustment	Vertical size expansion level for text images to compensate cut letter. Letters might be cut on the edges in case of the EAST model accuracy problems. That parameter defines how much vertical size should be expanded comparing to the original height	0
rotation_angle_threshold	For detections with angled text boxes node applies rotation to display text vertically. Parameters allows disabling rotation for angles below this value.	0