Method, system and device for multi-label object detection based on an object detection network

Patent metadata

Field	Value
Publication number	US-11429818-B2
Application number	US-201916960072-A
Country	US
Kind code	B2
Filing date	Dec 10, 2019
Priority date	Mar 7, 2019
Publication date	Aug 30, 2022
Grant date	Aug 30, 2022

Abstract

A multi-label object detection method based on an object detection network includes: selecting an image of an object to be detected as an input image; based on a trained object detection network, obtaining a class of the object to be detected, coordinates of a center of the object to be detected, and a length and a width of a detection rectangular box according to the input image; and outputting the class of the object to be detected, the coordinates of the center of the object to be detected, and the length and the width of the detection rectangular box. The method of the present invention can perform real-time and accurate object detection on different classes of objects with improved detection speed and accuracy, and can solve the problem of object overlapping and occlusion during the object detection.

First claim

What is claimed is:

1. A multi-label object detection method based on an object detection network, comprising:
step S 10 , selecting an image of an object to be detected as an input image;
step S 20 , based on the object detection network, obtaining a class of the object to be detected, coordinates of a center of the object to be detected, and a length and a width of a detection rectangular box according to the input image; and
step S 30 , outputting the class of the object to be detected, the coordinates of the center of the object to be detected, and the length and the width of the detection rectangular box;
wherein the object detection network is trained and obtained by replacing a low-resolution feature layer in a You Only Look Once-V3 (YOLO-V3) network with a densely connected convolutional network.

2. The multi-label object detection method based on the object detection network according to claim 1 , wherein, a method of training the object detection network comprises the following steps:
step B 10 , adjusting an attribute of each image in an obtained training image set according to a standard format to obtain a standardized training image set;
step B 20 , detecting a batch of images in the standardized training image set by using the object detection network, and calculating a training error of each classifier of the object detection network;
step B 30 , when a preset number of training iterations is not reached or the training error is greater than or equal to a preset threshold, obtaining a variation of a parameter of each layer in the object detection network and updating a parameter of the object detection network by an error back propagation method; and
step B 40 , detecting the standardized training image set in a batching sequence after the parameter of the object detection network is updated, and iteratively updating the parameter of the object detection network by the error back propagation method in step B 30 until the preset number of the training iterations is reached or the training error is lower than the preset threshold to obtain the object detection network.

3. The multi-label object detection method based on the object detection network according to claim 2 , wherein, the training error is calculated by the following formula:
Loss=Error coord Error iou +Error cls

where, Loss denotes the training error, Error coord denotes a prediction error of the coordinates, Error iou denotes an Intersection over Union (IoU) error between a predicted bounding box and a true bounding box, and Error cls denotes a classification error.

4. The multi-label object detection method based on the object detection network according to claim 3 , wherein, the prediction error of the coordinates is calculated by the following formula:

Error
coord

=

λ
coord

⁢

∑

i
=
1

S
2

⁢

∑

j
=
0

B

⁢

l
ij
obj

⁡

[

(

x
i

-

x
^

i

)

2

+

(

y
i

-

y
^

i

)

2

]

+

λ
coord

⁢

∑

i
=
1

S
2

⁢

∑

j
=
0

B

⁢

l
ij
obj

[

(

w
i

-

w
^

i

)

2

+

(

h

Assignees

• Inst Automation Cas

Inventors

• Yang Guodong
• Tian Yunong
• Li En
• Liang Zize
• Tan Min
• Jing Fengshui
• Gao Zishu
• Wang Hao
• Sun Yuansong
• Lu Sixi

Classifications

• G06N3/084 Primary

  Backpropagation, e.g. using gradient descent · CPC title

• G06V10/7796

  based on specific statistical tests · CPC title

• G06V10/774

  Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting · CPC title

• G06V10/82

  using neural networks · CPC title

• G06V10/25

  Determination of region of interest [ROI] or a volume of interest [VOI] · CPC title