CN-122023773-A - Non-maximum value inhibition method applied to image guidance head target detection

CN122023773ACN 122023773 ACN122023773 ACN 122023773ACN-122023773-A

Abstract

The invention discloses a non-maximum value inhibition method applied to image guidance head target detection, which is used for repositioning an optimal prediction frame generated by Yolov-tiny target detection algorithm, then using DIOU to replace IOU as a judgment criterion for inhibiting redundant prediction frames when inhibiting other prediction frames, using Gaussian penalty function to reduce the confidence coefficient of the redundant prediction frames, effectively improving the positioning accuracy of the prediction frames, improving the omission problem in dense scenes, improving the target detection accuracy and having stronger engineering application value.

Inventors

ZHOU WENJIE
LIU JIEYING
LIU ZHEN
LI LEPING
YANG WEIPING
LU XIN
ZHOU WEI
HUANG PENG
ZHANG LINGLING
YANG YONGFU
YANG YONGDA

Assignees

湖南华南光电(集团)有限责任公司

Dates

Publication Date: 20260512
Application Date: 20260130

Claims (5)

1. A non-maximum value suppression method applied to image guidance head target detection is characterized in that input information is set as a prediction frame set The output information is an optimal prediction frame set F, and the implementation steps of the method are as follows: Step one, inputting an image with resolution of W multiplied by H into Yolov-tiny target detection algorithm; Step two, three tensors output by Yolov-tiny algorithm are obtained, the three tensors correspond to the prediction frame information of the small, medium and large objects respectively, and the three output tensors are analyzed to obtain target information; Step three, arranging the candidate frames in the prediction frame set B according to the confidence scores of different categories from high to low, and taking the candidate frame with the highest confidence score of the category For the optimal prediction frame, will Removing from the candidate box set B; step four, calculating a current optimal prediction frame Candidate frames of the same class as candidate frame set B Find the proximity P of (2) less than the threshold Is a prediction frame of the model; step five, calculating a current optimal prediction frame And (3) repositioning the optimal prediction frame by using the average offset between the optimal prediction frame and the prediction frame obtained in the step (II) Obtaining a new optimal prediction frame Will be Adding the optimal prediction frame set F; Step six, utilizing a new optimal prediction frame Suppressing other prediction frames in the prediction frame set B; and step seven, recursively executing the steps three to six until the prediction frame set B is empty.
2. The non-maximum suppression method for image guidance head target detection according to claim 1, wherein the calculation method for analyzing three output tensors to obtain target information in the second step is as follows: wherein x, y is the coordinate value of the prediction frame in the input image, w, h is the width and height of the prediction frame in the input image, And For predicting the coordinates of a value within a corresponding cell, And Is the width and the height of the anchor frame, 、、、 Respectively the offset of the prediction frames, Is the width and height of the unit cell.
3. The non-maximum suppression method for image guidance head target detection according to claim 1, wherein in the fourth step, the current optimal prediction frame Prediction frames of the same class as candidate frame set B The method of calculating the proximity P of (c) is as follows: Wherein the dot is For optimal prediction frame The lower left corner of the body is provided with a left-hand corner, For predicting frames Lower left corner, m is the optimal prediction frame Upper right corner, n is the prediction frame At the upper right-hand corner of the frame, Representation points Sum point The Manhattan distance between the two is calculated as follows: and transforming the coordinate range to be between 0 and 1 by adopting the standardization of the coordinate points of the prediction frame, wherein the calculation method of the standardization of the coordinates is as follows: 。
4. The non-maximum suppression method for image guidance head target detection according to claim 1, wherein in the fifth step, the optimal prediction frame is The calculation method of the average offset between the prediction frames meeting the requirement of the second step is as follows: Wherein, the For the prediction box in set B meeting the second requirement, As a threshold value for the proximity it is, For average offset, optimal prediction block Adding an average offset Obtaining the final optimal prediction frame 。
5. An application to an image as claimed in claim 1a non-maximum value suppression method for pilot head target detection, the method is characterized in that in the step six, a new optimal prediction frame is utilized Suppressing other predicted frames in the predicted frame set B, which are the optimal predicted frames calculated by the fourth step And in prediction frame set B When two prediction frames are close to or overlap with each other, if the center point distance between the two prediction frames is smaller, judging that the two frames belong to the same target, and the calculation method is as follows: wherein A is an optimal prediction frame D is the area in the prediction frame set B Is d is the area of And The Euclidean distance between the center points of (c) is And Diagonal length of the minimum bounding rectangle; For the following The prediction frames larger than the threshold value apply Gaussian penalty, and the calculation method is as follows: Wherein, the For prediction in frame set B Is used to determine the confidence level of the (c) in the (c), And Y is a DIOU threshold value.

Description

Non-maximum value inhibition method applied to image guidance head target detection Technical Field The invention belongs to the field of detection of an image target of a seeker, and particularly relates to a non-maximum value inhibition method applied to detection of the image target of the seeker. Background Object detection is an important task in the field of computer vision that involves identifying one or more objects in an image or video frame and determining the location of each object. Current image-guided head target detection algorithms generate a large number of prediction frames of different positions and sizes during the prediction phase, most of these prediction frames are gathered in the region that may contain the target of interest, there is a large amount of redundancy, and people typically use non-maximal suppression algorithms to remove redundant prediction frames, preserving the optimal prediction frames. The traditional non-maximum suppression algorithm generally selects the prediction frame with the highest confidence in the same class as the optimal prediction frame, so that the correlation between the class confidence and the positioning accuracy is often ignored, and the high class confidence does not mean that the positioning accuracy of the frame is high. In addition, when the traditional non-maximum suppression algorithm suppresses other prediction frames by using the prediction frames with high confidence, the unconditional frames are directly removed, but in a scene with dense target distribution, different targets may be blocked, when two targets are relatively close, the frames with high confidence are often removed due to high overlapping degree, and the problem of missed detection in the dense scene is easily caused. Disclosure of Invention Aiming at the problems that the optimal prediction frame selected in the frame selection stage cannot accurately position the target and the problem of missed detection easily occurring when other prediction frames are restrained in the traditional non-maximum suppression algorithm and the improved algorithm thereof, the invention provides a non-maximum suppression method applied to image guidance head target detection to improve the target detection precision. The invention adopts the following technical scheme that the non-maximum value inhibition method applied to image guidance head target detection sets input information as a prediction frame setThe output information is an optimal prediction frame set F, and the implementation steps of the method are as follows: Step one, inputting an image with resolution of W multiplied by H into Yolov-tiny target detection algorithm. And step two, three tensors output by Yolov-tiny algorithm are obtained, the three tensors respectively correspond to the prediction frame information of the small, medium and large objects, and the three output tensors are analyzed to obtain target information. Step three, arranging the candidate frames in the prediction frame set B according to the confidence scores of different categories from high to low, and taking the candidate frame with the highest confidence score of the categoryFor the optimal prediction frame, willRemoved from candidate box set B. Step four, calculating a current optimal prediction frameCandidate frames of the same class as candidate frame set BFind the proximity P of (2) less than the thresholdIs included in the prediction block. Step five, calculating a current optimal prediction frameAnd (3) repositioning the optimal prediction frame by using the average offset between the optimal prediction frame and the prediction frame obtained in the step (II)Obtaining a new optimal prediction frameWill beAnd adding the optimal prediction frame set F. Step six, utilizing a new optimal prediction frameOther prediction blocks in the prediction block set B are suppressed. And step seven, recursively executing the steps three to six until the prediction frame set B is empty. Further, the calculation method for analyzing the three output tensors to obtain the target information in the second step is as follows: wherein x, y is the coordinate value of the prediction frame in the input image, w, h is the width and height of the prediction frame in the input image, AndFor predicting the coordinates of a value within a corresponding cell,AndIs the width and the height of the anchor frame,、、、Respectively the offset of the prediction frames,Is the width and height of the unit cell. Further, in the fourth step, the current optimal prediction framePrediction frames of the same class as candidate frame set BThe method of calculating the proximity P of (c) is as follows: Wherein the dot is For optimal prediction frameThe lower left corner of the body is provided with a left-hand corner,For predicting framesLower left corner, m is the optimal prediction frameUpper right corner, n is the prediction frameAt the upper right-hand corner of the frame,Representation