CN113947723A - High-resolution remote sensing scene target detection method based on size balance FCOS - Google Patents

High-resolution remote sensing scene target detection method based on size balance FCOS Download PDF

Info

Publication number
CN113947723A
CN113947723A CN202111143539.9A CN202111143539A CN113947723A CN 113947723 A CN113947723 A CN 113947723A CN 202111143539 A CN202111143539 A CN 202111143539A CN 113947723 A CN113947723 A CN 113947723A
Authority
CN
China
Prior art keywords
target
positive sample
centrality
anchor frame
remote sensing
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202111143539.9A
Other languages
Chinese (zh)
Other versions
CN113947723B (en
Inventor
尹建伟
陈振乾
尚永衡
蔡钰祥
杨莹春
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Zhejiang University ZJU
Original Assignee
Zhejiang University ZJU
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Zhejiang University ZJU filed Critical Zhejiang University ZJU
Priority to CN202111143539.9A priority Critical patent/CN113947723B/en
Publication of CN113947723A publication Critical patent/CN113947723A/en
Application granted granted Critical
Publication of CN113947723B publication Critical patent/CN113947723B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computational Linguistics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Evolutionary Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a high-resolution remote sensing scene target detection method based on a size balance FCOS (fuzzy C-means OS). according to the centrality and border regression stage of a size balance coefficient in a target detection module of the FCOS, a centrality coefficient is dynamically adjusted according to regression information of each target, reasonable weight is distributed to the border regression process of each positive sample, a high-resolution remote sensing target is used for detecting a remote sensing data set to carry out model training, and a model remote sensing ground object is used for identifying. The method fully considers the defects of targets with different sizes under the FCOS centrality evaluation system, performs loss weight reinforcement on samples with positive samples distributed at the edges in a small target anchor frame, suppresses the redundant loss contribution in a large target, and realizes target size balance; the size balance FCOS improves the target detection precision in a high-resolution remote sensing scene under the condition of not introducing extra overhead to a model reasoning stage.

Description

High-resolution remote sensing scene target detection method based on size balance FCOS
Technical Field
The invention belongs to the technical field of computer vision and remote sensing image application, and particularly relates to a high-resolution remote sensing scene target detection method based on a size balance FCOS.
Background
In recent years, the satellite technology has been rapidly developed, the application field of remote sensing images is continuously expanded, and the remote sensing image system plays a great role in the fields of meteorology, geology, agriculture and forestry, military, smart cities and the like. By remote sensing detection, multi-level, visual angle and time observation can be carried out on the images of a large area on the earth in a short time, and the method is an important means for acquiring environmental information and earth resources; through the target detection technology based on deep learning, the extraction of the ground features of the high-resolution remote sensing image can be efficiently and accurately realized, and the manual interpretation cost is reduced.
The remote sensing target has the characteristics of size proportion, various rotation directions, numerous small targets and the like, and the method creates a great test for the traditional target detection algorithm based on an anchor-based (anchor-based). The anchor-free target detection network gets rid of the disadvantages of anchor-based network such as preset fixed proportion and size of anchor frame prior knowledge, and realizes target detection by directly regression based on the distance from a central point to a target frame. The FCOS (full probabilistic One-Stage Object Detection) target Detection network takes a point falling into a target as a positive sample to carry out regression through the centrality, and gives a higher border regression loss weight to the positive sample falling into the target centre through the centrality, and the edge is a low weight; for some small targets with only a small number of positive samples, the weight distribution strategy has a high probability that the positive samples fall on the target edge, and the loss contribution of the positive samples is reduced due to low weight distribution; for large targets, more positive samples are allocated, which causes a contribution of redundancy loss, thereby affecting the overall target detection quality.
A target detection technology, namely, the Faster R-CNN based on an anchor frame is provided in the documents [ Ren S, He K, Girshick R, et al. fast R-CNN: war real-time object detection with region processing networks [ J ]. Advances in neural information processing systems,2015,28: 91-99 ], and the target object is detected by setting the anchor frame with fixed proportion and size through prior knowledge. However, the target size change ratio in the remote sensing scene is large, the rotation directions are various, and the small targets are numerous, so that a great test is created for the target detection technology based on the anchor frame.
A non-anchor frame target detection technology FCOS is proposed in a document [ Tian Z, Shen C, Chen H, et al.Fcos: full volumetric one-stage object detection [ C ]// Proceedings of the IEEE/CVF international conference on computer vision.2019: 9627-9636 ], regression detection is carried out on a target based on a positive sample point, geometric and spatial distribution characteristics of a remote sensing ground object target are adapted, and the defects based on preset proportion and size in the anchor frame detection technology are overcome.
The existing FCOS gives different weights to different positive sample points in a centrality mode, but the centrality cannot correctly reflect the regression quality of an anchor frame, particularly in the case that a remote sensing scene mainly comprises a small target. For small targets, the weight distribution strategy of centrality has a larger probability, so that the positive sample points of the small targets fall on the edge of the anchor frame of the small targets, and the positive sample points acquire low weight mistakenly; for large targets, redundant positive samples exist, which contribute to low-quality regression loss, and influence final target detection.
Disclosure of Invention
In view of the above, the invention provides a high-resolution remote sensing scene target detection method based on a size balance FCOS, which adopts different processing strategies for positive samples in targets with different scales, so that a small target regresses to provide sufficient loss contribution, and simultaneously, redundant positive samples of a large target are inhibited, so that targets with different sizes can obtain ideal detection effects, thereby achieving size balance, and finally realizing the improvement of target detection precision on the premise of not introducing extra overhead.
A high-resolution remote sensing scene target detection method based on a size balance FCOS comprises the following steps:
(1) extracting the centrality and the GIoU (Generalized Intersection over Union) of a positive sample in the image, wherein the positive sample is a pixel point falling in a target anchor frame (artificially marked);
(2) counting the mean value mu of the centrality of all positive samples in the imagecAnd standard deviation σcAnd further calculates the centrality threshold ct(ii) a Counting the number of positive samples falling into each target anchor frame in the image, and obtaining the average distribution number mu of the positive samples of the target anchor frame by averagingpa；
(3) According to the number of positive samples in the target anchor frame and the average number mu of positive samples distributed in the target anchor framepaCalculating the weighting coefficient w of the target anchor frameg；
(4) The mean value mu of all positive samples GIoU in the statistical imagegAnd standard deviation σgFurther, the GIoU threshold value t is calculatedgSimultaneously finding out the maximum centrality of the positive sample in the target anchor frame;
(5) calculating and determining a size balance coefficient of each positive sample in the image;
(6) and constructing a size balance FCOS model, adjusting and designing a loss function L of the model, inputting image characteristics into the model, and training the model by using a gradient descent method according to the loss function L so as to perform target detection on ground objects in the remote sensing image.
Further, the centrality threshold c is calculated in the step (2) by the following formulat；
ct＝μc+λcσc
Wherein: lambda [ alpha ]cThe weight parameter is set and is used for adjusting the influence of the centrality standard deviation on the centrality threshold.
Further, in the step (3), the weighting coefficient w of the target anchor frame is calculated by the following formulag；
Figure BDA0003284899820000031
Wherein: n is a radical ofgThe number of positive samples in the target anchor frame.
Further, the GIoU threshold value t is calculated in the step (4) by the following equationg；
tg＝μg+λgσg
Wherein: lambda [ alpha ]gIs a set weight parameter for adjusting the effect of the GIoU standard deviation on the GIoU threshold.
Further, the specific implementation manner of the step (5) is as follows: for any positive sample in the image, judging the maximum centrality c of the positive sample in the target anchor frame where the positive sample is locatedg,maxAnd a centrality threshold ctThe size relationship of (1): if c isg,max<ctCalculating the size balance coefficient q of the positive sample by the following relational expression;
q＝max(ct,cw)
cw＝wgcm
wherein: c. CmIs the centrality of the positive sample;
if c isg,max≥ctAnd the GIoU of the positive sample is smaller than the GIoU threshold value tgIf the size balance coefficient q of the positive sample is 0;
if c isg,max≥ctAnd the GIoU of the positive sample is greater than or equal to the GIoU threshold tgLet the size balance coefficient q of the positive sample be cm。
Further, the expression of the loss function L in the step (6) is as follows:
Figure BDA0003284899820000041
wherein: n is a radical ofposThe number of positive samples in the image, (x, y) the coordinates of a certain positive sample in the corresponding feature pyramid layer,
Figure BDA0003284899820000042
when in use
Figure BDA0003284899820000043
Then
Figure BDA0003284899820000044
Otherwise
Figure BDA0003284899820000045
Figure BDA0003284899820000046
Size balance coefficient, L, for coordinate (x, y) corresponding to positive samplecls() As a function of Focal distance for classification,
Figure BDA0003284899820000047
Figure BDA0003284899820000048
when in use
Figure BDA0003284899820000049
And is
Figure BDA00032848998200000410
Then
Figure BDA00032848998200000411
Otherwise
Figure BDA00032848998200000412
px,yAnd
Figure BDA00032848998200000413
respectively representing the prediction class label probability vector of the positive sample corresponding to the coordinate (x, y) and the corresponding truth label, Lreg() Is the GIoU loss function for bounding box regression, tx,yAnd
Figure BDA00032848998200000414
for coordinates (x, y) corresponding to the distance vectors of the positive samples from the prediction anchor frame and the target anchor frame, respectively, Lctr() As a cross-entropy function, qx,yFor coordinates (x, y) corresponding to the predicted centrality of the positive sample, λ1And λ2Are all given weight parametersAnd (4) counting.
According to the method, the centrality coefficient is dynamically adjusted according to regression information of each target by utilizing the centrality and border regression stages of the size balance coefficient in a target detection module of the FCOS, reasonable weight is distributed to the border regression process of each positive sample, a remote sensing data set is detected by using a high-resolution remote sensing target to carry out model training, and the model is used for identifying remote sensing ground objects. The method fully considers the defects of targets with different sizes under the FCOS centrality evaluation system, performs loss weight reinforcement on samples with positive samples distributed at the edges in a small target anchor frame, suppresses the redundant loss contribution in a large target, and realizes target size balance; the size balance FCOS improves the target detection precision in a high-resolution remote sensing scene under the condition of not introducing extra overhead to a model reasoning stage.
Drawings
FIG. 1 is a schematic diagram of a network architecture for a size balanced FCOS employed in the present invention.
FIG. 2 is a schematic flow chart of the high-resolution remote sensing scene target detection method of the invention.
Fig. 3(a) -3 (f) are schematic diagrams illustrating distribution and processing procedures of positive samples in different sizes of targets, where fig. 3(a) is a case where an anchor frame of a small target and a positive sample thereof fall at an edge of the target anchor frame, fig. 3(b) is a case where the anchor frame of the positive sample within the anchor frame of the small target regresses, fig. 3(c) is a case where the positive sample falls within the distribution of the center of the small target after calculation by a size balance coefficient, fig. 3(d) is a case where the anchor frame of the large target and the positive sample thereof fall at the edge of the target anchor frame, fig. 3(e) is a case where the anchor frames of all the positive samples within the large target regresses, and fig. 3(f) is a case where the positive sample with low quality is removed by the size balance coefficient.
Detailed Description
In order to more specifically describe the present invention, the following detailed description is provided for the technical solution of the present invention with reference to the accompanying drawings and the specific embodiments.
The invention relates to a high-resolution remote sensing scene target detection method based on a size balance FCOS, which comprises the following steps:
(1) and in the size balance FCOS training stage, acquiring the centrality and the GIoU of the RoI.
The FCOS extracts features through a backbone network, and then allocates targets with different scales to different pyramid network layers for down-sampling by using a feature pyramid network; and the target detection module of the FCOS regresses the characteristics of each layer pixel by pixel, screens out the positive sample falling in the target anchor frame, obtains the regression probability, the centrality and the distance between the coordinate of the positive sample and the four edges of the target anchor frame, and calculates the GIoU value according to the distance.
The size balance FCOS network structure adopted by the invention is shown in figure 1, an input image firstly passes through a feature extraction network ResNet to obtain features of C3-C5 with different resolutions, and then a feature pyramid network FPN is used to obtain P3-P7 features with different resolutions.
The target detection module of the size balance FCOS extracts different layer characteristics on the FPN, and two groups of convolutions are used for completing a target detection task; each set of convolutions comprises 4 convolution kernels of 256 channels of 3 x 3, the corresponding pyramid network characteristic of the input is
Figure BDA0003284899820000051
l is the corresponding pyramid layer number, HlAnd WlIs corresponding to PlThe resolution of (a); finally, the features obtained through the convolution group pass through convolution kernels of 1 3 multiplied by 3T channel, and finally the prediction category of the target is obtained
Figure BDA0003284899820000052
Where T is the total number of classes of the object and Cls represents the probability of each class for each point of the current feature layer.
The regression of the anchor frame and the centrality shares a group of convolution characteristics, and 1 convolution of 4 channels and 1 channel of 3 multiplied by 3 is respectively adopted for the characteristics obtained by the convolution group to obtain
Figure BDA0003284899820000053
Predicted anchor frame and predicted centrality of
Figure BDA0003284899820000054
Where 4 channels represent a channel where t ═ is (l, r, t,b) representing the distances from the four sides of the target anchor frame of the current point prediction, and the centrality prediction represents the centrality coefficient of the current sample point.
And (3) taking the samples falling into the target anchor frame as positive samples to carry out regression, and calculating the centrality and the GIoU value of the positive samples by the following formulas:
Figure BDA0003284899820000061
Figure BDA0003284899820000062
in the formula: t is t*＝(l*,r*,t*,b*) Is the distance vector from the positive sample point to the left, right, top and bottom four edges of the target anchor frame, a is the anchor frame of the positive sample prediction, B is the anchor frame of the target, and C is the minimum convex hull that can contain both a and B.
(2) Global target information perception: counting the mean value and the standard deviation of the centrality of all positive samples in the current training image, and calculating a centrality threshold value according to the mean value and the standard deviation of the centrality; and calculating the number of positive samples falling into each target anchor frame, and counting the number of positive samples obtained by average distribution of each target anchor frame.
The work flow of global target information perception is shown in the first column on the left side of fig. 2, and the mean value μ of centrality of all positive samples is calculated firstcStandard deviation σcCalculating the centrality threshold c according to the centrality mean and the standard deviationt＝μc+λcσcWherein λ iscThe super parameter is a super parameter of the centrality threshold and is used for adjusting the influence of the centrality standard deviation on the centrality threshold. And then, calculating the number of positive samples distributed to each target anchor frame, and for the positive samples falling into a plurality of target anchor frames, classifying the positive samples into the target anchor frame with the smallest area. Calculating the number of evenly distributed positive samples of each target anchor frame
Figure BDA0003284899820000063
Wherein N ispIs the total number of positive samples, N, of all layers of the pyramid networkAIs the total number of target anchor frames, μpaIs the average of the number of positive samples assigned to each target anchor block.
In FIGS. 3(a) and 3(d), the total number of targets NaNumber of samples N ═ 5pAverage number of positive samples, μ, assigned to each targetpa2.5. The positive sample centrality in the two target anchor frames is Cg1＝[0.2]，Cg2＝[0.8,0.7,0.5,0.3]Mean centrality μc0.5, standard deviation of centrality σc0.23, according to the results of table 1, λ is set herec0.25, centrality threshold ct＝μc+λcσc＝0.56。
TABLE 1
λc -0.50 -0.25 0 0.25 0.50
mAP 37.6 38.0 37.8 37.7 37.6
The mAP is an average precision mean (mean average precision) of the size balance FCOS and is used for evaluating the target detection precision, and ResNet-50 and FPN are used as feature extraction networks in evaluation models.
(3) And (3) carrying out regression quality statistics on the target anchor frame positive sample: calculating a size balance coefficient according to the number of positive samples distributed in the current target anchor frame and the average value of the positive samples distributed by the global target; calculating a GIoU mean value and a standard deviation of a positive sample in the current target, and calculating a GIoU threshold value according to the GIoU mean value and the standard deviation; the maximum centrality of all positive samples in the target is calculated.
The workflow of the regression quality statistics of the positive samples of the target anchor frame is shown in the middle column of fig. 2, in this step, firstly, the target set G is traversed, and all the positive sample sets M in the current target anchor frame G are processedgCalculate MgMaximum centrality c of the positive sampleg,max(ii) a Calculating MgGIoU mean μ of all positive samplesgStandard deviation σg. Calculating M according to the mean value and standard deviation of GIoUgThe GIoU threshold of (1) is tg＝μg+λgσgWherein λ isgIs a super parameter of the GIoU threshold and is used to adjust the effect of the GIoU standard deviation on the threshold.
Calculating weighting coefficients
Figure BDA0003284899820000071
Wherein N isgIs the number of positive samples in the current target anchor frame g by dividing NgIs placed in the denominator such that wgInversely proportional to the number of falling positive samples, so that smaller targets get a higher weight addition.
In FIGS. 3(b) and 3(e), the GIoU values of the samples in the two target anchor boxes are Dg1＝[0.8]，Dg2＝[0.9,0.8,0.5,0.2](ii) a The maximum centrality of the samples in the two target anchor frames is cmax,g1＝0.2，cmax,g20.8; according to Ng1＝1，Ng2＝4，μpa＝2.5，
Figure BDA0003284899820000072
Calculating to obtain a size balance coefficient wg1＝1.58，wg2＝0.79。
Calculate each target Anchor frame DgMean value of middle GIoUg1＝0.8，μg20.6, standard deviation σg1＝0，σg20.27; according to the results of Table 2, λ is set hereg-0.25, according to the GIoU threshold calculation formula tg＝μg+λgσgTo obtain tg1＝0.8，tg2＝0.53。
TABLE 2
λg -0.50 -0.25 0 0.25 0.50
mAP 38.0 37.9 38.2 38.1 37.8
(4) And setting a size balance coefficient, and determining the centrality coefficient of the current positive sample to carry out weighting, removal or retention according to the GIoU and the centrality value of the target where the current positive sample is located. HeadFirst, go through MgThe weighted center degree c of the positive sample m is obtainedw＝wgcmWherein c ismIs the centrality of the current positive sample m. Then, the maximum centrality c of the positive sample in the current target anchor frame is judgedg,maxAnd a centrality threshold ctThe size relationship of (1): if c isg,maxLess than centrality threshold ctThis case is considered to be a positive sample within the extreme target anchor box. At this time, the centering degree needs to be weighted, and the size balance coefficient q of the target is max (c)t,cw) I.e. the large values of the centrality threshold and the weighted centrality, ensure that the small targets whose positive samples all fall on the edge of the target anchor frame are weight protected. For targets greater than the threshold, if the GIoU value d of the positive samplemIf the value is less than the GIoU threshold value, the low-quality regression is considered, the size balance coefficient q is set to be 0, otherwise the original centrality c of the q is keptm。
As shown in the first column on the right in FIG. 2 and in FIG. 3(c), for the target g1, since c isg,max<ctWhere the positive samples regress to obtain a high quality anchor frame, but the centrality is suppressed, where the centrality is weighted, cw1＝wg1cm10.3, the positive sample centrality q in the target is set to max (c)t,cw) 0.56; the centrality distribution of FIG. 3(c) is a distribution curve of the centrality threshold by selecting max (c)t,cw) Ensuring that the sample is near the target center point.
For target anchor frame g2, as shown in FIG. 3(e), due to the maximum centration cg2,max≥ctAt this time, the regression of the anchor frame with high weight and quality can be ensured, and redundant positive samples in the target need to be removed; traversal of positive samples within g2, GIoU value D for g2 positive samplesg2,30.5 and Dg2,40.5, since two GIoU values are less than tg20.53, belonging to the low-quality anchor frame prediction, the size balance coefficient q is set to 0, and the positive sample is removed, and the original centrality of the remaining positive sample is retained, as shown in fig. 3(f), the removed positive sample is represented by "x", and the final remaining prediction anchor frame is retained by the solid rectangle.
(5) And during the calculation of the loss function, multiplying the obtained size balance coefficient by the frame regression loss of each positive sample, and enabling the centrality branch of the FCOS to carry out regression on the size balance coefficient to complete the training stage.
The computational expression of the loss function is as follows:
Figure BDA0003284899820000081
in the formula: n is a radical ofposThe number of positive samples in the image, (x, y) the coordinates of a certain positive sample in the corresponding feature pyramid layer,
Figure BDA0003284899820000082
when in use
Figure BDA0003284899820000083
Then
Figure BDA0003284899820000084
Otherwise
Figure BDA0003284899820000085
Figure BDA0003284899820000086
Size balance coefficient, L, for coordinate (x, y) corresponding to positive samplecls() As a function of Focal distance for classification,
Figure BDA0003284899820000087
Figure BDA0003284899820000088
when in use
Figure BDA0003284899820000089
And is
Figure BDA00032848998200000810
Then
Figure BDA00032848998200000811
Otherwise
Figure BDA00032848998200000812
px,yAnd
Figure BDA00032848998200000813
respectively representing the prediction class label probability vector of the positive sample corresponding to the coordinate (x, y) and the corresponding truth label, Lreg() Is the GIoU loss function for bounding box regression, tx,yAnd
Figure BDA00032848998200000814
for coordinates (x, y) corresponding to the distance vectors of the positive samples from the prediction anchor frame and the target anchor frame, respectively, Lctr() As a cross-entropy function, qx,yFor coordinates (x, y) corresponding to the predicted centrality of the positive sample, λ1And λ2Are given weight parameters.
Table 3 shows the comparison of FCOS and size-balanced FCOS for other positive sample processing strategies, where no centrality is introduced in FCOS and ATSS removes only below the GIoU threshold t compared to size-balanced FCOSgThe suppressed small target is not weighted; the size balance coefficient realizes the improvement of target detection precision under the condition of not causing extra overhead, and has more superiority compared with other balance coefficient methods.
TABLE 3
Method Without centrality FCOS ATSS Size balanced FCOS
mAP 37.5 37.8 38.0 38.2
(6) And performing target detection on the ground objects in the input preprocessed remote sensing image by using the trained size balance FCOS model.
The foregoing description of the embodiments is provided to enable one of ordinary skill in the art to make and use the invention, and it is to be understood that other modifications of the embodiments, and the generic principles defined herein may be applied to other embodiments without the use of inventive faculty, as will be readily apparent to those skilled in the art. Therefore, the present invention is not limited to the above embodiments, and those skilled in the art should make improvements and modifications to the present invention based on the disclosure of the present invention within the protection scope of the present invention.

Claims (6)

1. A high-resolution remote sensing scene target detection method based on a size balance FCOS comprises the following steps:
(1) extracting the centrality and the GIoU of a positive sample in the image, wherein the positive sample is a pixel point falling in a target anchor frame;
(2) counting the mean value mu of the centrality of all positive samples in the imagecAnd standard deviation σcAnd further calculates the centrality threshold ct(ii) a Counting the number of positive samples falling into each target anchor frame in the image, and obtaining the average distribution number mu of the positive samples of the target anchor frame by averagingpa；
(3) According to the number of positive samples in the target anchor frame and the average number mu of positive samples distributed in the target anchor framepaCalculating the weighting coefficient w of the target anchor frameg；
(4) Counting places in imagesMean value μ of GIoU with positive samplegAnd standard deviation σgFurther, the GIoU threshold value t is calculatedgSimultaneously finding out the maximum centrality of the positive sample in the target anchor frame;
(5) calculating and determining a size balance coefficient of each positive sample in the image;
(6) and constructing a size balance FCOS model, adjusting and designing a loss function L of the model, inputting image characteristics into the model, and training the model by using a gradient descent method according to the loss function L so as to perform target detection on ground objects in the remote sensing image.
2. The method for detecting the target in the high-resolution remote sensing scene according to claim 1, characterized in that: in the step (2), the centrality threshold c is calculated by the following formulat；
ct＝μc+λcσc
Wherein: lambda [ alpha ]cIs the set weight parameter.
3. The method for detecting the target in the high-resolution remote sensing scene according to claim 1, characterized in that: in the step (3), the weighting coefficient w of the target anchor frame is calculated by the following formulag；
Figure FDA0003284899810000011
Wherein: n is a radical ofgThe number of positive samples in the target anchor frame.
4. The method for detecting the target in the high-resolution remote sensing scene according to claim 1, characterized in that: the GIoU threshold value t is calculated in the step (4) by the following formulag；
tg＝μg+λgσg
Wherein: lambda [ alpha ]gIs the set weight parameter.
5. According to claim 1The high-resolution remote sensing scene target detection method is characterized by comprising the following steps: the specific implementation manner of the step (5) is as follows: for any positive sample in the image, judging the maximum centrality c of the positive sample in the target anchor frame where the positive sample is locatedg，maxAnd a centrality threshold ctThe size relationship of (1): if c isg，max＜ctCalculating the size balance coefficient q of the positive sample by the following relational expression;
q＝max(ct，cw)
cw＝wgcm
wherein: c. CmIs the centrality of the positive sample;
if c isg，max≥ctAnd the GIoU of the positive sample is smaller than the GIoU threshold value tgIf the size balance coefficient q of the positive sample is 0;
if c isg，max≥ctAnd the GIoU of the positive sample is greater than or equal to the GIoU threshold tgLet the size balance coefficient q of the positive sample be cm。
6. The method for detecting the target in the high-resolution remote sensing scene according to claim 1, characterized in that: the expression of the loss function L in the step (6) is as follows:
Figure FDA0003284899810000021
wherein: n is a radical ofposThe number of positive samples in the image, (x, y) the coordinates of a certain positive sample in the corresponding feature pyramid layer,
Figure FDA0003284899810000022
when in use
Figure FDA0003284899810000023
Then
Figure FDA0003284899810000024
Otherwise
Figure FDA0003284899810000025
Figure FDA0003284899810000026
Size balance coefficient, L, for coordinate (x, y) corresponding to positive samplecls() As a function of Focal distance for classification,
Figure FDA0003284899810000027
Figure FDA0003284899810000028
or 1, when
Figure FDA0003284899810000029
And is
Figure FDA00032848998100000210
Then
Figure FDA00032848998100000211
Otherwise
Figure FDA00032848998100000212
px，yAnd
Figure FDA00032848998100000213
respectively representing the prediction class label probability vector of the positive sample corresponding to the coordinate (x, y) and the corresponding truth label, Lreg() Is the GIoU loss function for bounding box regression, tx，yAnd
Figure FDA00032848998100000214
for coordinates (x, y) corresponding to the distance vectors of the positive samples from the prediction anchor frame and the target anchor frame, respectively, Lctr() As a cross-entropy function, qx，yFor coordinates (x, y) corresponding to the predicted centrality of the positive sample, λ1And λ2Are given weight parameters.
CN202111143539.9A 2021-09-28 2021-09-28 High-resolution remote sensing scene target detection method based on size balance FCOS Active CN113947723B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111143539.9A CN113947723B (en) 2021-09-28 2021-09-28 High-resolution remote sensing scene target detection method based on size balance FCOS

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111143539.9A CN113947723B (en) 2021-09-28 2021-09-28 High-resolution remote sensing scene target detection method based on size balance FCOS

Publications (2)

Publication Number Publication Date
CN113947723A true CN113947723A (en) 2022-01-18
CN113947723B CN113947723B (en) 2024-07-02

Family

ID=79328947

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111143539.9A Active CN113947723B (en) 2021-09-28 2021-09-28 High-resolution remote sensing scene target detection method based on size balance FCOS

Country Status (1)

Country Link
CN (1) CN113947723B (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117685881A (en) * 2024-01-31 2024-03-12 成都建工第七建筑工程有限公司 Sensing and detecting system for concrete structure entity position and size deviation

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111091105A (en) * 2019-12-23 2020-05-01 郑州轻工业大学 Remote sensing image target detection method based on new frame regression loss function
CN111583282A (en) * 2020-05-18 2020-08-25 联想(北京)有限公司 Image segmentation method, device, equipment and storage medium
WO2020173036A1 (en) * 2019-02-26 2020-09-03 博众精工科技股份有限公司 Localization method and system based on deep learning
CN112800955A (en) * 2021-01-27 2021-05-14 中国人民解放军战略支援部队信息工程大学 Remote sensing image rotating target detection method and system based on weighted bidirectional feature pyramid

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2020173036A1 (en) * 2019-02-26 2020-09-03 博众精工科技股份有限公司 Localization method and system based on deep learning
CN111091105A (en) * 2019-12-23 2020-05-01 郑州轻工业大学 Remote sensing image target detection method based on new frame regression loss function
CN111583282A (en) * 2020-05-18 2020-08-25 联想(北京)有限公司 Image segmentation method, device, equipment and storage medium
CN112800955A (en) * 2021-01-27 2021-05-14 中国人民解放军战略支援部队信息工程大学 Remote sensing image rotating target detection method and system based on weighted bidirectional feature pyramid

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
刘斌平;周越;: "一种新颖的无锚框三维目标检测器", 中国体视学与图像分析, no. 01, 25 March 2020 (2020-03-25) *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117685881A (en) * 2024-01-31 2024-03-12 成都建工第七建筑工程有限公司 Sensing and detecting system for concrete structure entity position and size deviation
CN117685881B (en) * 2024-01-31 2024-06-04 成都建工第七建筑工程有限公司 Sensing and detecting method for concrete structure entity position and size deviation

Also Published As

Publication number Publication date
CN113947723B (en) 2024-07-02

Similar Documents

Publication Publication Date Title
CN109614985B (en) Target detection method based on densely connected feature pyramid network
CN111091105B (en) Remote sensing image target detection method based on new frame regression loss function
CN108416307B (en) Method, device and equipment for detecting pavement cracks of aerial images
CN110287932B (en) Road blocking information extraction method based on deep learning image semantic segmentation
CN108038846A (en) Transmission line equipment image defect detection method and system based on multilayer convolutional neural networks
CN112836713A (en) Image anchor-frame-free detection-based mesoscale convection system identification and tracking method
CN108985250A (en) A kind of traffic scene analytic method based on multitask network
CN107633226B (en) Human body motion tracking feature processing method
CN112464911A (en) Improved YOLOv 3-tiny-based traffic sign detection and identification method
CN112101278A (en) Hotel point cloud classification method based on k nearest neighbor feature extraction and deep learning
CN111967480A (en) Multi-scale self-attention target detection method based on weight sharing
CN111079739B (en) Multi-scale attention feature detection method
CN112084869A (en) Compact quadrilateral representation-based building target detection method
CN106023257A (en) Target tracking method based on rotor UAV platform
CN111145145B (en) Image surface defect detection method based on MobileNet
CN111738114B (en) Vehicle target detection method based on anchor-free accurate sampling remote sensing image
CN113177456B (en) Remote sensing target detection method based on single-stage full convolution network and multi-feature fusion
CN110852317A (en) Small-scale target detection method based on weak edge
CN115527133A (en) High-resolution image background optimization method based on target density information
CN115995042A (en) Video SAR moving target detection method and device
CN113505712B (en) Sea surface oil spill detection method of convolutional neural network based on quasi-balance loss function
CN113947723B (en) High-resolution remote sensing scene target detection method based on size balance FCOS
CN112597875A (en) Multi-branch network anti-missing detection aerial photography target detection method
CN110796716B (en) Image coloring method based on multiple residual error network and regularized transfer learning
CN116363532A (en) Unmanned aerial vehicle image traffic target detection method based on attention mechanism and re-parameterization

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant