Object detection model generation method and electronic device and computer readable storage medium using the same

Patent metadata

Field	Value
Publication number	US-2022044438-A1
Application number	US-202117403902-A
Country	US
Kind code	A1
Filing date	Aug 17, 2021
Priority date	Aug 5, 2020
Publication date	Feb 10, 2022
Grant date	—

Abstract

An object detection model generation method as well as an electronic device and a computer readable storage medium using the same are provided. The method includes: during the iterative training of the to-be-trained object detection model, the detection accuracy of the iteration nodes of the object detection model is sequentially determined according to the node order, and the mis-detected negative samples of the object detection model at the iteration nodes with the detection accuracy less than or equal to a preset threshold are enhanced. Then the object detection model is trained at the iteration node based on the enhanced negative samples and a first amount of preset training samples. After the training at the iteration nodes are completed, it returns to the step of sequentially determining the detection accuracy of the iteration nodes of the object detection model until the training of the object detection model is completed.

First claim

What is claimed is:

1 . An object detection model generation method, comprising steps of:
inputting a preset first amount of training samples into a to-be-trained object detection model, and performing an iterative training on the object detection model, wherein the training samples include one or more positive samples and one or more negative samples; determining an untrained current iteration node as an object iteration node according to a node order, and obtaining one or more model parameters of the object detection model corresponding to a previous iteration node of the object iteration node; determining a detection accuracy of the object detection model at the object iteration node based on the one or more model parameters; obtaining one or more enhanced negative samples by enhancing one or more mis-detected negative samples of the object detection model at the object iteration node according to a preset negative sample enhancement rule, in response to the detection accuracy being less than or equal to a preset accuracy threshold; training the object iteration node based on the one or more enhanced negative samples and the preset first amount of the training samples; and returning to the determining the untrained current iteration node as the object iteration node according to the node order after the object iteration node is trained until the object detection model is trained.

2 . The method of claim 1 , wherein after the determining the detection accuracy of the object detection model at the object iteration node based on the one or more model parameters, further comprises steps of:
training the object detection model at the object iteration node based on the preset first amount of the training samples, in response to the detection accuracy being larger than the preset accuracy threshold.

3 . The method of claim I, wherein the obtaining the one or more enhanced negative samples by enhancing the one or more mis-detected negative samples of the object detection model at the object iteration node according to the preset negative sample enhancement rule comprises:
obtaining the one or more mis-detected negative samples of the object detection model at the object iteration node; obtaining one or more spliced images by splicing a preset second amount of the mis-detected negative samples and the positive samples at intervals; and obtaining the one or more enhanced negative samples by cropping all the spliced images according to a preset cropping rule.

4 . The method of claim 3 , wherein the spliced image is a grid image, and the mis-detected negative samples and the one or more positive samples are placed in grids of the grid image at intervals.

5 . The method of claim 4 , wherein the obtaining the one or more enhanced negative samples by cropping all the spliced images according to the preset cropping rule comprises:
cropping the grid image according to a preset cropping size to obtain the grids of the grid image: and taking the grid including the one or more mis-detected negative samples as the one or more enhanced negative samples.

6 . The method of claim 3 , wherein the obtaining the one or more mis-detected negative samples of the object detection model at the object iteration node comprises:
determining the negative sample having detected as the positive sample when the object detection model performs object detection on each of the preset first amount of the training samples at the previous iteration node; and obtaining the one or more mis-detected negative samples of the object detection model at the object iteration node by obtaining all the negative samples having detected as the positive sample.

7 . The method of claim 1 , wherein the determining the detection accuracy of the object detection model at the object iteration node based on the one or more model parameters comprises:
taking the one or more model parameters as one or more variable of the object detection model at the object iteration node, and determining an Intersection over Union of the object detection model performing object detection on each of a preset third amount of the training samples; and determining the detection accuracy of the object detection model at the object iteration node according to each Intersection over Union.

8 . An electronic device, comprising:
a processor; a memory coupled to the processor; and one or more computer programs stored in the memory and executable on the processor; wherein, the one or more computer programs comprise: instructions for inputting a preset first amount of training samples into a to-be-trained object detection model, and performing an iterative training on the object detection model, wherein the training samples include one or more positive samples and one or more negative samples; instructions for determining an untrained current iteration node as an object iteration node according to a node order, and obtaining one or more model parameters of the object detection model corresponding to a previous iteration node of the object iteration node; instructions for determining a detection accuracy of the object detection model at the object iteration node based on the one or more model parameters; instructions for obtaining one or more enhanced negative samples by enhancing one or more mis-detected negative samples of the object detection model at the object iteration node according to a preset negative sample enhancement rule, in response to the detection accuracy being less than or equal to a preset accuracy threshold; instructions for training the object iteration node based on the one or more enhanced negative samples and the preset first amount of the training samples; and instructions for returning to the determining the untrained current iteration node as the object iteration node according to the node order after the object iteration node is trained until the object detection model is trained.

9 . The electronic device of claim 8 , wherein the one or more computer programs further comprise:
instructions for training the object detection model at the object iteration node based on the preset first amount of the training samples, in response to the detection accuracy being larger than the preset accuracy threshold.

10 . The electronic device of claim K, wherein the instructions for obtaining the one or more enhanced negative samples by enhancing the one or more mis-detected negative samples of the object detection model at the object iteration node according to the preset negative sample enhancement rule comprise:
instructions for obtaining the one or more mis-detected negative samples of the object detection model at the object iteration node; instructions for obtaining one or more spliced images by splicing a preset second amount of the mis-detected negative samples and the positive samples at intervals; and instructions for obtaining the one or more enhanced negative samples by cropping all the spliced images according to a preset cropping rule.

11 . The electronic device of claim 10 , wherein the spliced image is a grid image, and the mis-detected negative samples and the one or more positive samples are placed in grids of the grid image at intervals.

12 . The electronic device of claim 11 , wherein the instructions tor obtaining the one or more enhanced negative samples by cropping all the spliced images according to the preset cropping rule comprise:
instructions for cropping the grid image according to a preset cropping size to obtain the grids of the grid image; and instructions for taking the grid including the one or more mis-detected negative samples as the one or more enhanced negat

Assignees

• Ubtech Robotics Corp Ltd

Inventors

• Liu Yepeng
• Zeng Yusheng
• Cheng Jun
• Gu Jing
• Wang Yue
• Pang Jianxin

Classifications

• G06N3/045

  Combinations of networks · CPC title

• G06N3/0455

  Auto-encoder networks; Encoder-decoder networks · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N3/08

  Learning methods · CPC title