Looking for breakthrough ideas for innovation challenges? Try Patsnap Eureka!

Weak and small human body target detection method based on accurate scale matching

A technology for human targets and detection methods, applied in the fields of computer vision and machine learning, can solve the problems of SM algorithm approximate processing uncertainty, imprecision, scale mismatch, etc., to balance information loss, improve performance, and promote similarity Effect

Active Publication Date: 2020-11-17
UNIVERSITY OF CHINESE ACADEMY OF SCIENCES
View PDF2 Cites 7 Cited by
  • Summary
  • Abstract
  • Description
  • Claims
  • Application Information

AI Technical Summary

Problems solved by technology

Since there may be many labeled objects with multiple scales in the image, the approximation of the SM algorithm is uncertain and inaccurate, resulting in the appearance of scale mismatch
like figure 1 As shown, the SM algorithm can roughly match the scale distribution of COCO with the scale distribution of TinyPerson, but there is a misalignment problem (as shown by the dotted rectangle)

Method used

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
View more

Image

Smart Image Click on the blue labels to locate them in the text.
Viewing Examples
Smart Image
  • Weak and small human body target detection method based on accurate scale matching
  • Weak and small human body target detection method based on accurate scale matching
  • Weak and small human body target detection method based on accurate scale matching

Examples

Experimental program
Comparison scheme
Effect test

Embodiment approach

[0109] According to a preferred embodiment of the present invention, the probabilistic structure repair method includes the following steps:

[0110] Step i, obtain the random number of the background of each picture in the pre-training data set in step 1;

[0111] In step ii, set the probability threshold p, and determine whether to introduce additional background by comparing the random number with the probability threshold.

[0112] Among them, in step ii, if the random number of the background of the original picture is greater than the probability threshold p, a new image is sampled from the pre-training data set as the background; if the random number of the background of the original picture is less than or equal to the probability threshold p, the inpainting method is used to repair the background of the original picture .

[0113] Preferably, the probability threshold p ranges from 0 to 1, preferably 0.4.

[0114] In the present invention, the labels of the newly sa...

Embodiment 1

[0145] 1. Dataset

[0146] The method described in the present invention is carried out in COCO and TinyPerson dataset:

[0147] COCO involves 80 classes of objects. The present invention performs network pre-training on a combination of the original training set (including 80k images), the subset (35k images) of the original verification set (Trainval35k) and the 5k subset (Minival) of the verification image.

[0148] TinyPerson is a tiny object detection dataset collected from high-quality videos and web images. In a total of 1,610 images, TinyPerson contains 72,651 low-resolution annotated human objects. 60% and 40% of the images in the dataset are randomly selected as training and testing sets. In TinyPerson, most annotation objects are smaller than 20x20 pixels in size. Subgraphs cut from the origin image are used as input during training and inference prediction. Finally, the results of the same image are combined using the NMS strategy.

[0149] 2. Implementation ...

experiment example 1

[0156] For the scale alignment between the pre-training data set and the target data set, the SM+ method (instance-level scale matching) described in the present invention and the SM method (image-level scale matching) described in the prior art are respectively used for detection, The result is as Figure 4shown.

[0157] Depend on Figure 4 It can be seen that after COCO is adjusted using the SM+ method of the present invention, its aligned distribution is closer to the distribution of TinyPerson. The SM+ method achieves more accurate scale matching, which can achieve higher detection performance.

[0158] Further, the Jensen-Shannon divergence was used to quantitatively measure the similarity between the scale distribution and the target distribution aligned by the RSM+ method, the MSM+ method of the present invention, and the RSM and MAM methods in the prior art.

[0159] Among them, the Jensen-Shannon divergence is a deformation of the Kullback-Leibler divergence, whic...

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to View More

PUM

No PUM Login to View More

Abstract

The invention discloses a weak and small human body target detection method based on accurate scale matching, and the method comprises the steps: aligning the scale distribution of a pre-training dataset and a target data set at an example level, and generating a more effective and more proper matching data set; in order to reduce the structural loss of the image caused by example-level alignmentdistribution, adopting a background processing method based on a probability structure and dynamically restoring an image by inhibiting image blurring and maintaining the context consistency around ahole. According to the method disclosed by the invention, the similarity between the pre-training data set and the target data set is effectively promoted, the information loss between an image structure and semantics can be effectively balanced, the performance on TinyPerson is obviously improved, and the performance is obviously improved in the aspects of average precision (AP) and missing rate(MR).

Description

technical field [0001] The invention relates to the technical fields of computer vision and machine learning, in particular to a method for detecting tiny objects in large-scale images, and in particular to a method for detecting weak and small human objects based on accurate scale matching. Background technique [0002] Human detection is an important topic in the field of computer vision, which has a wide range of applications such as surveillance, driving assistance, and rapid rescue at sea. With the rapid development of data-driven deep convolutional neural networks (CNNs), detector research has made significant progress. However, the detector performs poorly when detecting tiny objects with fewer pixels (e.g., smaller than 20x20 pixels in size), such as traffic signs, human objects in aerial photography, and so on. [0003] In order to make better use of CNN-based detectors, some human-annotated datasets for human detection are proposed and made public. However, datas...

Claims

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to View More

Application Information

Patent Timeline
no application Login to View More
Patent Type & Authority Applications(China)
IPC IPC(8): G06K9/62G06K9/00G06T5/00
CPCG06V40/10G06V2201/07G06F18/22G06F18/241G06F18/214G06T5/00G06T5/90
Inventor 韩振军蒋楠余学辉陈鹏飞宫宇琦韩许盟彭潇珂王岿然吴狄黄智勋焦建彬叶齐祥万方
Owner UNIVERSITY OF CHINESE ACADEMY OF SCIENCES
Who we serve
  • R&D Engineer
  • R&D Manager
  • IP Professional
Why Patsnap Eureka
  • Industry Leading Data Capabilities
  • Powerful AI technology
  • Patent DNA Extraction
Social media
Patsnap Eureka Blog
Learn More
PatSnap group products

Browse by: Latest US Patents, China's latest patents, Technical Efficacy Thesaurus, Application Domain, Technology Topic, Popular Technical Reports.

© 2024 PatSnap. All rights reserved.Legal|Privacy policy|Modern Slavery Act Transparency Statement|Sitemap|About US| Contact US: help@patsnap.com