Interactive small sample semantic segmentation training method

A technology of semantic segmentation and training methods, applied in the direction of character and pattern recognition, instruments, computer components, etc., can solve the problems of limited application and promotion, and achieve the effect of solving the large workload of labeling data

Pending Publication Date: 2022-01-14
深圳市玻尔智造科技有限公司
View PDF0 Cites 0 Cited by
  • Summary
  • Abstract
  • Description
  • Claims
  • Application Information

AI Technical Summary

Problems solved by technology

However, the existing target detection tasks are trained based on a large number of labeled images, which limits the application and promotion in some scenarios.

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
  • Interactive small sample semantic segmentation training method
  • Interactive small sample semantic segmentation training method
  • Interactive small sample semantic segmentation training method

Examples

Experimental program
Comparison scheme
Effect test

Embodiment 1

[0034] Embodiment 1 of this application provides such as figure 1 The training method for the interactive small-sample semantic segmentation shown:

[0035] Step 1, such as image 3 As shown, let the user draw a series of points of different categories on the original picture, and use the points to mark the target to be learned (abnormal sample 1, normal sample 2); compared with the traditional target detection, the labeling of the present invention does not need to be very detailed , you only need to draw a series of points of different categories on the image to be labeled to achieve high-precision detection;

[0036] Step 2. Input the picture marked with the target to be learned in step 1 into the pre-training model, perform feature extraction, and obtain a feature map;

[0037] Step 3. Project the marked points onto the feature map to obtain feature vectors and category labels, and use the logistic regression model to classify and predict the original marked points;

[...

Embodiment 2

[0042] On the basis of Embodiment 1, Embodiment 2 of the present application provides the following figure 2 Specific steps for step 2 shown:

[0043] Step 2.1. The picture marked with the target to be learned is pooled through the conv1 layer and the maxpool layer to obtain a feature map;

[0044] Step 2.2. After the feature map is processed by layer1, input layer2 to extract features, and then go through layer3 for further feature extraction; improve the expressive ability of the network, and further extract deep-level features;

[0045] Step 2.3: Interpolate the output obtained from layer 1 with the output obtained from layer 2, and then concatenate with the output obtained from layer 3 to obtain multiple spliced ​​feature maps of different scales as the final output, which is more conducive to the network Learn and express.

Embodiment 3

[0047] On the basis of Embodiment 1, Embodiment 3 of the present application provides the specific steps of Step 3:

[0048] Step 3.1, point the marked point i (x,y) is projected onto the feature map to get the marked point on the feature map i (x new ,y new ):

[0049] point i (x new ,y new ) = point i (x,y)*(w t / w,h t / h)) (1)

[0050] In the above formula, (x, y) represents the corresponding position coordinates of the original label point; (x new ,y new ) represents the position coordinates of the original label point projected on the feature map; w t 、h t are the width and height of the feature map respectively; w and h are the width and height of the original map respectively;

[0051] Step 3.2, obtain the feature vector T and the category label by projecting the marked points on the feature map;

[0052] Step 3.3, use the logistic regression model to classify and predict the original label points:

[0053]

[0054] In the above formula, x represents th...

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 relates to an interactive small sample semantic segmentation training method, which comprises the following steps of: drawing points of different categories on an original picture, and marking a to-be-learned target by using the points; inputting the picture marked with the to-be-learned target into a pre-training model, and performing feature extraction; projecting the marking points to the feature map to obtain feature vectors and category labels, and classifying and predicting the original marking points by using a logistic regression model; and using a single picture as a small sample, and learning features of different types of marking points of the small sample. The interactive small sample semantic segmentation training method has the advantages that the method can achieve training of a good model effect only through one picture through the interactive marking of a learning target, the feature extraction and training, and the over-detection and missing detection repairing of a sample picture, and the problems that in a current common target detection method, the data labeling workload is large, and the generalization ability is weak when a small sample data training model is used are solved.

Description

technical field [0001] The invention belongs to the field of computer vision target detection, and in particular relates to a training method for interactive semantic segmentation of small samples. Background technique [0002] The target detection task is one of the basic tasks of computer vision, the main task is to classify and locate the target in the image. However, the existing target detection tasks are trained based on a large number of labeled images, which limits the application and promotion in certain scenarios. By applying a semi-supervised method with less labeled data or a weakly supervised method with incompletely matched labeled data, it is more important to use very little labeled data to learn a model with a certain generalization ability, which is also required for small-sample learning. solved problem. Contents of the invention [0003] The purpose of the present invention is to overcome the deficiencies in the prior art and provide an interactive sm...

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
IPC IPC(8): G06V10/764G06V10/774G06V10/778G06K9/62
CPCG06F18/217G06F18/214G06F18/2415
Inventor 杨培文张成英梁惠莹于振东张辽
Owner 深圳市玻尔智造科技有限公司
Who we serve
  • R&D Engineer
  • R&D Manager
  • IP Professional
Why Eureka
  • Industry Leading Data Capabilities
  • Powerful AI technology
  • Patent DNA Extraction
Social media
Try Eureka
PatSnap group products

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

© 2024 PatSnap. All rights reserved.Legal|Privacy policy|Modern Slavery Act Transparency Statement|Sitemap