CN109255320B

CN109255320B - Improved non-maximum suppression method

Info

Publication number: CN109255320B
Application number: CN201811018713.5A
Authority: CN
Inventors: 李宏亮; 廖加竞; 孙旭; 刘玮; 何慕威
Original assignee: University of Electronic Science and Technology of China
Current assignee: University of Electronic Science and Technology of China
Priority date: 2018-09-03
Filing date: 2018-09-03
Publication date: 2020-09-25
Anticipated expiration: 2038-09-03
Also published as: CN109255320A

Abstract

The invention discloses an improved non-maximum suppression method. The invention carries out inhibition treatment by a mode of multiple iterative screening: sorting the rectangular frames to be processed according to the confidence coefficient from high to low, reserving the rectangular frame with the highest current confidence coefficient, sequentially traversing whether the proportion of the sum of the overlapping areas of the remaining rectangular frames and all the reserved rectangular frames to the whole image area is greater than a threshold value, and if so, deleting the rectangular frames; and then, the rectangular frames which are not deleted and not reserved are used as a new round of rectangular frames to be processed to continue the suppression processing until only one rectangular frame to be processed exists. Through the method, the technical problems that the traditional non-maximum value inhibition method is small in quantity of rectangular frames output by irregular targets, high in overlapping rate and low in fitting degree of the rectangular frames to the targets are solved.

Description

Improved non-maximum suppression method

Technical Field

The invention belongs to the technical field of target detection, and particularly relates to an improved non-maximum suppression method in deep learning target detection.

Background

In deep learning target detection, non-maximum suppression processing is generally performed on an output result (rectangular frame) to remove redundant repetitive frames. However, the conventional non-maximum suppression method has the following 3 problems when used for detecting irregular targets:

(1) in each comparison, the currently selected box in the conventional method is compared with only one box with the highest confidence level. The method has the advantages of reducing the number of rectangular frames and obtaining local maximum values for the rule target. However, in the detection of irregular targets, the number of rectangular frames obtained by such a method is too small to sufficiently fit the shape of the target object;

(2) if the threshold of the traditional method is just reduced, although the number of output rectangular frames is increased, the side effect of high overlapping rate of the rectangular frames is caused;

(3) for irregular targets, a plurality of rectangular frames with large size difference are often needed to fit the shapes of the irregular targets, and when the irregular targets are compared with a threshold value in a traditional method, the proportion of the overlapping area of the two frames to the total area of the two frames is calculated. Such a method is greatly affected by the area of the rectangular frame. The smaller rectangular frame is small in area and can be easily deleted in the method.

Disclosure of Invention

The invention aims to: aiming at the technical problems of the traditional non-maximum suppression method that the number of rectangular frames output by an irregular target is small, the overlapping rate is high and the fitting degree of the rectangular frames to the target is low, an improved non-maximum suppression method is provided.

The improved non-maxima suppression method of the present invention comprises the steps of:

step S1: sorting all the rectangular frames to be processed, which are to be subjected to non-maximum suppression, from high to low according to confidence coefficients to obtain a first set T1;

step S2: placing the first element in the first set T1 into a second set T2, wherein the initial value of the set T2 is an empty set;

step S3: judging whether the number of the elements in the first set T1 is 1, if not, deleting the first element in the first set T1, and then executing the step S4; if yes, go to step S8;

step S4: sequentially traversing all the rectangular boxes in the first set T1;

step S5: judging whether the traversal is finished, if so, jumping to step S2; otherwise, executing step S6;

step S6: judging whether the ratio of the sum of the overlapping areas of the traversed current rectangular frame and all the rectangular frames in the second set T2 to the area of the whole graph is larger than a preset threshold (the preferable value range is 0.01-0.03), if so, executing a step S7; otherwise, jumping to step S4;

wherein, the full map area refers to the full map area of the image to be detected where the rectangular frame to be processed is located.

Deeply learning the area of an input image of a target detection network;

step S7: deleting the traversed current rectangular box from the set T1, and jumping to step S4;

step S8: the union of the sets T1 and T2 is output.

In summary, due to the adoption of the technical scheme, compared with the prior art, the invention has the beneficial effects that:

(1) by adjusting the threshold in step S6, the number of output rectangular frames after the suppression processing can be increased, that is, if the threshold is increased, the number of output rectangular frames can be increased;

(2) in step S6, the currently traversed rectangular frame and all the retained rectangular frames are compared, whereas in the conventional non-maximum suppression method, the current rectangular frame is only compared with one rectangular frame with the highest confidence, and the improvement method makes the overlapping rate between the output rectangular frames lower;

(3) in step S6, the ratio of the total overlapping area of the current rectangular frame and all the reserved rectangular frames to the total area of the whole image is calculated, while the ratio of the overlapping area of the two rectangular frames to the total area of the two rectangular frames is calculated by the conventional non-maximum suppression method.

Drawings

FIG. 1 is a process flow diagram of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention will be described in further detail with reference to the following embodiments and accompanying drawings.

Referring to fig. 1, the specific processing steps of the present invention are as follows:

step S1: acquiring all rectangular frames output by a deep learning target detection network, namely all rectangular frames to be subjected to non-maximum suppression;

step S2: sorting all the rectangular frames from high to low according to the confidence coefficient to obtain a first set T1;

step S3: the first element (the rectangle box with the highest current confidence) in the first set T1 is placed into the second set T2 (the initial value is an empty set);

step S4: judging whether the rectangular box with the highest current confidence coefficient is the last element of the first set T1 (namely judging whether the number of the elements of the first set T1 is 1), if not, deleting the first element (the rectangular box with the highest current confidence coefficient) in the first set T1, and then executing the step S5; if yes, go to step S9;

step S5: sequentially traversing all the rectangular boxes in the first set T1;

step S6: judging whether the traversal is finished, if so, jumping to step S3; otherwise, executing step S7;

step S7: judging whether the ratio of the sum of the overlapping areas of the traversed current rectangular frame and all the rectangular frames in the second set T2 to the area of the full image (the image to be detected input into the deep learning target detection network) is greater than 0.02, if so, executing step S8; otherwise, jumping to step S5;

step S8: deleting the traversed current rectangular box from the set T1 (i.e., deleting the rectangular box whose proportion of the sum of the overlapping areas of all the rectangular boxes in the second set T2 to the total map area is greater than 0.02), and jumping to step S5;

step S9: all undeleted rectangular boxes are output, i.e., the union of sets T1 and T2 is output.

While the invention has been described with reference to specific embodiments, any feature disclosed in this specification may be replaced by alternative features serving the same, equivalent or similar purpose, unless expressly stated otherwise; all of the disclosed features, or all of the method or process steps, may be combined in any combination, except mutually exclusive features and/or steps.

Claims

1. An improved non-maxima suppression method for deep learning target detection, comprising the steps of:

step S1: acquiring all output results output by the deep learning target detection network, taking each rectangular frame corresponding to the output results as a rectangular frame to be processed for non-maximum suppression, and sequencing all rectangular frames to be processed from high to low according to confidence coefficients to obtain a first set T1;

step S4: sequentially traversing all the rectangular boxes in the first set T1;

step S6: judging whether the ratio of the sum of the overlapping areas of the traversed current rectangular frame and all the rectangular frames in the second set T2 to the total image area is greater than a preset threshold, if so, executing a step S7; otherwise, jumping to step S4;

wherein, the whole image area refers to the area of the input image of the deep learning target detection network;

step S8: the union of the sets T1 and T2 is output.

2. The method according to claim 1, wherein in step S6, the preferred value range of the preset threshold is: 0.01 to 0.03.

3. The method of claim 2, wherein in step S6, the threshold is preferably set to 0.02.