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

A Human-Machine Collaborative Method for Semi-automatic Labeling of Image Object Detection Data

A target detection and human-machine collaboration technology, applied in the field of target detection, can solve the problems of reduced labeling accuracy, time-consuming, heavy workload, etc., and achieve the effect of improving running speed, excellent performance, and improving labeling fault tolerance

Active Publication Date: 2021-05-04
TIANJIN UNIV
View PDF11 Cites 0 Cited by
  • Summary
  • Abstract
  • Description
  • Claims
  • Application Information

AI Technical Summary

Problems solved by technology

In the case of a large number of pictures and dense targets, the workload of this kind of labeling is huge. After labeling for a long time or labeling small targets, the labeling accuracy will decrease, and when there are too many categories to be selected Labeling is also time consuming

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
  • A Human-Machine Collaborative Method for Semi-automatic Labeling of Image Object Detection Data
  • A Human-Machine Collaborative Method for Semi-automatic Labeling of Image Object Detection Data
  • A Human-Machine Collaborative Method for Semi-automatic Labeling of Image Object Detection Data

Examples

Experimental program
Comparison scheme
Effect test

Embodiment 1

[0032] An embodiment of the present invention provides a human-machine collaborative target detection data labeling method, the method includes the following steps:

[0033] 101: Improved target detection model Cascade R-CNN [5] (cascaded region-based convolutional neural network), remove RPN [1] (Regional candidate area network), continue to use the cascaded sub-network structure to achieve multiple corrections to the bounding box, and introduce dynamic reasoning to ensure running speed;

[0034] Furthermore, since the candidate area is directly provided by the user at this time, the RPN is no longer needed to extract the candidate area, and the RPN simplified network model is removed here, and then the cascaded network is added. This method cascades three sub-networks with the same structure for regressing bounding boxes at the back of the model. The proposals input by the latter two sub-networks are the bounding boxes output by the previous sub-network. Finally, a dynami...

Embodiment 2

[0040] The scheme in embodiment 1 is further introduced below in conjunction with specific examples and calculation formulas, see the following description for details:

[0041] 1. Data preparation

[0042] The present invention uses the general target detection dataset COCO2017 for training, which is released by Microsoft and contains more than 100,000 pictures, which can be used for multiple tasks such as target detection and semantic segmentation. Among them, the target detection task contains 80 categories of different scales and shapes.

[0043] 2. Improvement of the model

[0044] For an input picture I, this HMC R-CNN adopts 3 cascaded RoI Heads in Cascade R-CNN (the sub-network used to classify and return candidate regions in Faster R-CNN, well known to those skilled in the art ) structure, the three RoI Heads use the function g 1 , g 2 , g 3 To indicate that the backbone network (backbone) used to extract image features is recorded as a function f, and the candidat...

Embodiment 3

[0093] The experimental result 1 that the embodiment of the present invention adopts is as figure 1 and figure 2 As shown, these two pictures respectively reflect the IoU distribution between the candidate area and the final output and the ground truth bounding box, showing that the improved model of this method can effectively correct the pseudo-random candidate area, and the IoU of the candidate area after the model is processed The distribution is clearly clustered towards the part with higher IoU.

[0094] The experimental results 2 used in the embodiment of the present invention are shown in Table 1. This result shows the results of HMC R-CNN trained using the training scheme described in this method under the test conditions described in this method. Compared with the results of Cascade R-CNN on the test set of COCO2017, the performance has been very significant. improvement. Especially the detection improvement for smaller targets is particularly obvious, with more ...

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 human-machine collaborative image target detection data semi-automatic labeling method. The method includes the following steps: obtaining an improved target detection model, that is, removing the region candidate area network and continuing to use the cascade sub-network structure of Cascade RCNN to realize the target detection model. Multiple corrections of the bounding box, while introducing a dynamic reasoning mechanism, judge the accuracy of the labeling results of the current sample according to the category score, and dynamically determine the number of sub-networks used; use the training set to train the improved target detection model, and use the verification set Test the target detection model, the candidate area is obtained by adding random scaling and offset to the bounding box in the ground truth; the user provides the candidate area to the trained model, and the model returns the corrected bounding box and target category as the labeling result. The invention assists the user in labeling, reduces the labeling burden, and improves the labeling error tolerance rate; the user supervises the labeling results, corrects a few wrong results, and uses it for fine-tuning the model.

Description

technical field [0001] The invention relates to the field of target detection, in particular to a semi-automatic tagging method for human-machine collaborative image target detection data. Background technique [0002] Object detection is a long-term, basic and challenging problem in computer vision. People have been studying in this field for decades. Its definition is as follows: For a given picture, target detection is to judge whether the picture is There is an instance of a certain category that has been trained. If it exists, return the horizontal position of the object and the category it belongs to. [0003] In the past few years when deep learning is popular, object detection has made breakthrough progress. In the past, due to the limitation of algorithm performance, there were very few categories that could be detected (such as pedestrian detection). The deep convolutional neural network (DCNN) can automatically learn the characteristics of the data, which makes ...

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 Patents(China)
IPC IPC(8): G06K9/32G06N3/04G06N3/08
CPCG06N3/08G06V10/255G06V2201/07G06N3/045
Inventor 朱鹏飞刘家旭汪廉杰胡清华
Owner TIANJIN UNIV
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