Target positioning method, system and computer equipment based on hierarchical class activation graph

A target positioning and hierarchical technology, applied in the field of deep learning, can solve problems such as inaccurate positioning ability, and achieve the effect of saving time and cost, reducing computational complexity, and achieving significant target positioning ability.

Active Publication Date: 2021-11-12
CHONGQING UNIV OF POSTS & TELECOMM
View PDF8 Cites 0 Cited by
  • Summary
  • Abstract
  • Description
  • Claims
  • Application Information

AI Technical Summary

Problems solved by technology

[0006] Based on the problems existing in the prior art, the present invention makes some modifications to the basic convolutional network to address the shortcoming of inaccurate positioning capabilities caused by the lack of underlying information in target positioning

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
  • Target positioning method, system and computer equipment based on hierarchical class activation graph
  • Target positioning method, system and computer equipment based on hierarchical class activation graph
  • Target positioning method, system and computer equipment based on hierarchical class activation graph

Examples

Experimental program
Comparison scheme
Effect test

Embodiment 1

[0047] A target positioning method based on a hierarchical class activation map of the present invention, the method includes inputting the image to be predicted into a convolutional hierarchical structure, extracting the hierarchical features of the image to be predicted, and generating a hierarchical class activation map of the image to be predicted ;Reserve some values ​​in the hierarchical class activation map, and generate a bounding box that can predict the target to be tested;

[0048] Among them, such as figure 1 As shown, the generation of the hierarchical class activation map includes the following steps:

[0049] S1, Construct the convolutional hierarchical structure of the image to be predicted, including adding a layer custom convolutional layer;

[0050] S2, setting the step size and padding of the custom convolution layer added in S1;

[0051] S3. Superimpose the custom convolution layers corresponding to the convolution layer 4-3 and the convolution layer 4-...

Embodiment 2

[0057] This embodiment presents another embodiment of the present invention. In this embodiment, input the image to be tested into the model, calculate the loss function, and train the model until the loss function converges. Otherwise, use the gradient descent algorithm to update each parameter and continue to input Go to the model for training; when the model is trained, input the image to be tested, and extract the feature maps of convolutional layer 4-3, convolutional layer 4-4, convolutional layer 5-3, and convolutional layer 5-4, According to the formula (5), the saliency map I of the classification is determined A and I B ; and these two significant maps are superimposed to obtain a hierarchical class activation map; retain some values ​​in the activation map, in this embodiment, select a value greater than 20% of the maximum activation value to retain; use it to generate a predicted bounding box.

[0058] Among them, the loss function may adopt several types of loss f...

Embodiment 4

[0076] This embodiment provides relevant descriptions of the target positioning system in the present invention;

[0077] The present invention also proposes a target positioning system based on a hierarchical class activation map, the system comprising:

[0078] An image acquisition module, configured to acquire an image to be predicted;

[0079] The hierarchical feature extraction module is used to extract the hierarchical features in the image to be predicted;

[0080] A hierarchical class activation map building block, used to construct a hierarchical class activation map from hierarchical features;

[0081] The predicted bounding box calculation module is used to predict the bounding box of the target to be tested in the image to be predicted according to the hierarchical class activation map;

[0082] Wherein, the hierarchical feature extraction module includes a VGG19 network structure, a self-defined convolutional layer unit, an overlay layer unit, and a pooling laye...

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 present invention relates to the field of deep learning and object detection. The present invention discloses a target positioning method, system and computer equipment based on a hierarchical activation map. The method constructs a hierarchical model, and adopts a global average pooling layer or The pyramid pooling layer is used to replace the traditional fully connected layer to avoid losing image structure information in the fully connected layer. This method collects corresponding feature information in multiple convolutional layers in the lower layer to obtain hierarchical class activation maps. The hierarchical activation map of the present invention not only collects feature maps from the last layer, but collects them in multiple convolutional layers in the lower layers, thereby reducing the loss of lower-layer image information and improving image positioning capabilities.

Description

technical field [0001] The present invention relates to the fields of deep learning and object detection, and specifically uses deep learning technology to realize target positioning under object detection; specifically, it is a target positioning method based on a hierarchical class activation map. Background technique [0002] In recent years, with the rapid rise of deep learning technology, research on object detection in the image field has made important progress. Among them, the most popular object detection algorithm can be divided into two types: (1) two-step method, that is, first generate a series of sparse candidate frames through the CNN network structure, and then classify these candidate frames. (2) One-step method, similar to the SSD idea, uses different scales of aspect ratios to densely sample images at different positions of the image, uses CNN to extract features, and directly classifies them. Among them, the target positioning under object detection is m...

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): G06T7/73G06K9/62
CPCG06T7/73G06T2207/20081G06T2207/20084G06T2207/20016G06V2201/07G06F18/241
Inventor 李鸿健程卓曾祥燕段小林汪美琦
Owner CHONGQING UNIV OF POSTS & TELECOMM
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