sb_t_svl_dl_par_network Struct Reference

Deep Learning network parameters. More...

#include <sb.h>

Collaboration diagram for sb_t_svl_dl_par_network:

Data Fields
sb_t_network_type	type
	Network type. More...
sb_t_size	input_size
	Network input size. More...
sb_t_image_format	image_format
	Network input image format. More...
int	n_channels
	Network input channels. More...
float	features_multiplier
	Features multiplier factor of the network. More...
sb_t_network_freeze_mode	freeze_mode
	Freeze mode to apply on network parameters. More...

Detailed Description

Deep Learning network parameters.

Used only by Deep Cortex and Deep Surface projects.

Definition at line 10518 of file sb.h.

Field Documentation

features_multiplier

float sb_t_svl_dl_par_network::features_multiplier

Features multiplier factor of the network.

Percentage of the available features effectively used by the network. The value is applied to all the layers except the first and the last one. Values lower than SB_SVL_DL_FEATURES_MULTIPLIER_DEFAULT decrease training and detection time, but may reduce the discrimination capability of the network.

Default

The default value is SB_SVL_DL_FEATURES_MULTIPLIER_DEFAULT.

Definition at line 10574 of file sb.h.

freeze_mode

sb_t_network_freeze_mode sb_t_svl_dl_par_network::freeze_mode

Freeze mode to apply on network parameters.

It determines the part of the current SVL parameters configuration to be frozen before running next SVL. It may be used only in case of incremental SVL or when SVL starts from an existing pre-training configuration file (i.e. sb_t_svl_dl_par::pre_training_file points to a valid SB_PRE_TRAINING_EXT file).
It is set to SB_NETWORK_FREEZE_MODE_NONE .

Warning

Readable only.

Definition at line 10584 of file sb.h.

image_format

sb_t_image_format sb_t_svl_dl_par_network::image_format

Network input image format.

Image format of the image that network expecting to elaborate. If the value is different from image format of the source image to be elaborated, the latter will be converted to be equal to the network one.
It is set to SB_IMAGE_FORMAT_RGB888 .

Warning

Readable only.

Definition at line 10557 of file sb.h.

input_size

sb_t_size sb_t_svl_dl_par_network::input_size

Network input size.

It determines, at which resolution, images must be resized before to be elaborated by the network.
SB_NETWORK_TYPE_SDINET0 is the only network with variabile input size, while the other networks have fixed input sizes.

Warning

Readable only for Deep Cortex projects.

Definition at line 10548 of file sb.h.

n_channels

int sb_t_svl_dl_par_network::n_channels

Network input channels.

Number of input channels of the network. The value is directly related to sb_t_svl_dl_par_network::image_format.
It is set = 3.

Warning

Readable only.

Definition at line 10565 of file sb.h.

type

sb_t_network_type sb_t_svl_dl_par_network::type

Network type.

It specifies the Convolutional Neural Network (CNN) algorithm used by functions sb_svl_run and sb_project_detection.
Allowed values for Deep Cortex projects are:

• SB_NETWORK_TYPE_EFFICIENTNET_B0 , SB_NETWORK_TYPE_EFFICIENTNET_B1 , SB_NETWORK_TYPE_EFFICIENTNET_B2 : they all are different versions of the same base network EfficientNet in increasing order of performance, input resolution and computational complexity. EfficientNet is proved to perform well also on complex image classification tasks with limited training and inference time. More information about EfficientNet are available in the official paper of the 2019: "EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks" (https://arxiv.org/abs/1905.11946).
  
• SB_NETWORK_TYPE_ICNET0_64 , SB_NETWORK_TYPE_ICNET0_128 : they all are different versions of the the same base network ICNet (Image Classification Network) in increasing order of performance, input resolution and computational complexity. ICNet is a proprietary network designed by FaberVision to solve efficiently simple and medium complex image classification vision tasks.
  

Allowed values for Deep Surface projects are:

• SB_NETWORK_TYPE_SDINET0 : SDINet (Surface Defect Inspection Network) is a proprietary network designed by FaberVision to solve efficiently object/defect segmentation tasks. It works with the input resolutions set by the user. It means that images/tiles are all resized to the selected resolution before to be elaborated.
  In order to train on and detect small defects instances in images at higher resolution try to increase the number of tiles or decrease the scale.

Definition at line 10540 of file sb.h.

---

The documentation for this struct was generated from the following file:

• sb.h