Unlock instant, AI-driven research and patent intelligence for your innovation.

Pooling method and device, object detection method, device and system and computer readable medium

An object detection and pooling technology, applied in the field of image processing, can solve problems such as inability to use convolutional features, incomplete response, and impact on accuracy

Inactive Publication Date: 2018-03-06
BEIJING KUANGSHI TECH CO LTD +1
View PDF2 Cites 3 Cited by
  • Summary
  • Abstract
  • Description
  • Claims
  • Application Information

AI Technical Summary

Problems solved by technology

In this process, the use of low-level convolutional features is often ignored
Although some methods have begun to consider the use of low-level convolutional features, most similar methods can only use different levels of convolutional features in a lower level.
[0004] In the prior art, when the convolutional network is fed forward, the mainstream maximum pooling is generally used to process the feature map. This pooling method only retains the response of the maximum value in a certain area. When the information of the image conforms to a special When distributed, it is very likely that the distribution cannot be fully reflected, and the convergence is poor, and the convolution features in the lower levels cannot be used, thus affecting the accuracy of object detection.

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
  • Pooling method and device, object detection method, device and system and computer readable medium
  • Pooling method and device, object detection method, device and system and computer readable medium
  • Pooling method and device, object detection method, device and system and computer readable medium

Examples

Experimental program
Comparison scheme
Effect test

Embodiment 1

[0061] The embodiment of the present invention provides a pooling method, see figure 1 with figure 2 As shown, the method includes the following steps:

[0062] S11: Determine an interception threshold according to the feature value of the feature map of the target image.

[0063] The feature value of the target image feature map is the value in the feature matrix of the feature map obtained by the feature extraction process of the target image. As a preferred implementation, the determination of the interception threshold in this embodiment is as follows:

[0064] According to the average value and standard deviation of the feature value of the target image feature map, the interception threshold is determined.

[0065] Specifically, the average value μ and the standard deviation σ of the feature values ​​of the feature map of the target image are first calculated; and then the interception threshold A is calculated according to the formula A=μ+1.5σ.

[0066] The truncati...

Embodiment 2

[0079] An embodiment of the present invention provides an object detection method, see Figure 4 with Figure 5 As shown, this method can be applied to deep convolutional networks in large-scale image recognition, and can also be applied to deep residual learning in the field of image recognition, specifically including the following steps:

[0080] S21: Perform multi-layer convolution processing on the input image to obtain multiple feature maps.

[0081] When implementing it, first import the input image to be detected, and then perform n-layer convolution processing on the input image. Generally speaking, the minimum value of n is 3, that is, at least 3 layers of convolution processing are performed. After the input image is processed by multi-layer convolution, multiple feature maps are obtained, such as feature figure 1 ,feature figure 2 ... feature map n.

[0082] S22: Use the pooling method as described in Embodiment 1 to perform pooling processing on the feature m...

Embodiment 3

[0102] The embodiment of the present invention also provides a pooling device, see Figure 7 As shown, the device includes: a threshold determination module 31 , a comparison module 32 , a feature map acquisition module 33 , and an average pooling module 34 .

[0103] Wherein, the threshold determination module 31 is used to determine the interception threshold according to the feature value of the target image feature map; the comparison module 32 is used to compare the feature value of the target image feature map with the interception threshold; the feature map acquisition module 33 is used for The eigenvalues ​​less than the interception threshold in the target image feature map are set to 0, and the other eigenvalues ​​remain unchanged to obtain the intercepted target image feature map; the mean pooling module 34 is used to average the intercepted target image feature map Value pooling to obtain the pooled target image feature map.

[0104]In the pooling device provided ...

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 provides a pooling method and device, an object detection method, device and system and a computer readable medium, and relates to the technical field of image processing. The pooling method comprises the steps of determining an intercepting threshold according to a feature value of a target image feature map; enabling the feature value of the target image feature map to be comparedwith the intercepting threshold; setting feature values, which are less than the intercepting threshold, in the target image feature map to be 0, keeping the other feature values to be unchanged, andacquiring an intercepted target image feature map; and performing mean pooling on the intercepted target image feature map to obtain a pooled target image feature map. According to the method, intercepting and pooling processing can be performed on the target image feature map through the processes of intercepting threshold determination and mean pooling, and thus low-level convolution features can be combined into high-level convolution features, thereby performing deep integration on the low-level convolution features, and achieving a purpose of rational use for the low-level features.

Description

technical field [0001] The present invention relates to the technical field of image processing, in particular to a pooling method, an object detection method, a device, a system, and a computer-readable medium. Background technique [0002] General object detection (general object detection) is an important part of the field of computer vision, and it plays an important role in many applications of computer vision (such as unmanned driving, robotics, aerospace positioning and face detection and other related applications). [0003] At present, the main method of general object detection is: based on the strategy of deep learning, the features output by the deep convolutional layer are used to represent the characteristics of the detected object, and the characteristics of the detected object are further used for classification. In this process, the use of low-level convolution features is often ignored. Although some methods have begun to consider the use of low-level conv...

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): G06K9/46G06K9/62G06N3/04G06N3/08
CPCG06N3/08G06V10/443G06V10/757G06N3/045
Inventor 王志成俞刚
Owner BEIJING KUANGSHI TECH CO LTD