Custom node for general purpose image transformations

This custom node takes image with dynamic shape (color, width, height) as an input. It performs multiple operations and produces an output:

• resize to desired width and height
• layout change between NCHW and NHWC
• color ordering between BGR, RGB (3 color channels) and GRAY (1 color channel)
• change data value range per channel: [0;255], [0;1], [-1;1]

Important to note that this node uses OpenCV for processing so for good performance results prefers NHWC layout. In other cases conversion applies which reduces performance of this node.

NOTE Exemplary configuration files are available in onnx model with server preprocessing demo and config with single node.

Building custom node library

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

git clone https://github.com/openvinotoolkit/model_server && cd model_server/src/custom_nodes
make NODES=image_transformation

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

You can also select base OS between RH 8.10 (redhat) and Ubuntu 22.04 (ubuntu) by setting BASE_OS environment variable.

make BASE_OS=redhat NODES=image_transformation

Custom node inputs

Input name	Description	Shape	Precision
image	Input image in an array format. Only batch size 1 is supported. Resolution is dynamic (node takes any width and height) but should be greater than 0. 1 and 3 color channels are supported. Data might be either in NCHW or NHWC format.	1,C,H,W or 1,H,W,C (configurable via parameter)	FP32

Custom node outputs

Output name	Description	Shape	Precision
image	Returns image after transformation. Transformations are configurable via parameters.	1,C,H,W or 1,H,W,C (configurable via parameter)	FP32

Custom node parameters

Parameter	Description	Default	Required
target_image_width	Desired image width after transformation. If not specified, width will not be changed.
target_image_height	Desired image height after transformation. If not specified, height will not be changed.
original_image_color_order	Input image color order	BGR
target_image_color_order	Output image color order. If specified and differs from original_image_color_order, color order conversion will be performed	BGR
original_image_layout	Input image layout. This is required to determine image shape from input shape		✓
target_image_layout	Output image layout. If specified and differs from original_image_layout, layout conversion will be performed
scale	All values will be divided by this value. When scale_values is specified, this value is ignored. read more
scale_values	Scale values to be used for the input image per channel. Input data will be divided by those values. Values should be provided in the same order as output image color order. read more
mean_values	Mean values to be used for the input image per channel. Values will be subtracted from each input image data value. Values should be provided in the same order as output image color order. read more
debug	Defines if debug messages should be displayed	false

NOTE: Subtracting mean values is performed before division by scale values.