CN114972982A - Remote sensing image target detection sample unbalance processing method based on improved oversampling - Google Patents
Remote sensing image target detection sample unbalance processing method based on improved oversampling Download PDFInfo
- Publication number
- CN114972982A CN114972982A CN202210425413.9A CN202210425413A CN114972982A CN 114972982 A CN114972982 A CN 114972982A CN 202210425413 A CN202210425413 A CN 202210425413A CN 114972982 A CN114972982 A CN 114972982A
- Authority
- CN
- China
- Prior art keywords
- category
- remote sensing
- categories
- image
- images
- 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.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/24—Aligning, centring, orientation detection or correction of the image
- G06V10/242—Aligning, centring, orientation detection or correction of the image by image rotation, e.g. by 90 degrees
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/25—Determination of region of interest [ROI] or a volume of interest [VOI]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Multimedia (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Medical Informatics (AREA)
- Software Systems (AREA)
- Databases & Information Systems (AREA)
- Computing Systems (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Artificial Intelligence (AREA)
- Image Analysis (AREA)
Abstract
The invention provides a remote sensing image target detection class imbalance processing method based on improved oversampling, which comprises the steps of firstly counting the class quantity and various target quantities in a remote sensing data set; then judging whether the data set has the problem of unbalanced category, if not, ending, and if so, turning to the next step; secondly, expanding the category images with small target amount; and finally, obtaining a target expanded data set. The method can overcome the defect that the prior oversampling technology is easy to cause model overfitting, thereby solving the problem of unbalanced detection categories of the remote sensing image targets.
Description
Technical Field
The field relates to the field of remote sensing image processing, in particular to a method for processing the unbalance problem of target detection categories of remote sensing images.
Background
The unbalanced detection of the remote sensing image target type means that the number of targets in each type in the data set is unbalanced, which is particularly common in the remote sensing data set, mainly because the number difference of various ground objects in nature is large, such as the ratio of automobiles to airplanes in a city, and the ratio of the types even exceeds 1000: 1. It has been proved that the imbalance of categories adversely affects the accuracy of the machine learning and deep learning algorithms, which affects the convergence of the algorithms in the training phase and also affects the generalization of the algorithms in the testing phase.
In the remote sensing image target detection, the solution to the class imbalance problem generally includes the following three types: the first type is loss function adjustment, the loss function determines a weight adjustment mode in model training, a focus loss function or a weighting loss function is generally adopted for a class unbalanced data set, and the algorithm does not directly solve the class imbalance problem, so that improvement of the model performance is limited. The second type is a sampling method, which generally includes oversampling and downsampling, where oversampling refers to performing multiple random sampling on a class image with a small target amount, and downsampling refers to removing a part of the class image with a large target amount. The over-sampling algorithm is generally considered to cause the class over-fitting of the model to a small amount of targets, mainly because the current over-sampling method is a simple copy of images containing the class, so that the model is over-fitted to the class of targets, and the generalization performance is reduced. Down-sampling would discard a large number of images, which would be detrimental to model convergence. The third category is sample weighting, which is essentially a combination of over-sampling and down-sampling, i.e., down-sampling the more targeted class images and over-sampling the less targeted class images, which generally reduces the total number of data set images to seek a quantitative balance between the classes, resulting in model overfitting.
Disclosure of Invention
The invention aims to provide a remote sensing image target detection class imbalance processing method based on improved oversampling, so as to improve the defects of the existing remote sensing image target detection class imbalance oversampling method.
In order to solve the technical problems, the technical scheme provided by the invention is as follows:
a remote sensing image target detection sample unbalance processing method based on improved oversampling comprises the following steps:
s1, counting the number N of categories and the target number [ N ] of each category in the remote sensing data set 1 ，N 2 ，N 3 ，...，N n ]；
S2, judging whether there is the problem of unbalanced classification, if there is unbalanced classification, going to step S3, otherwise, ending the flow;
s3, expanding the category images with small target amount;
s4: and obtaining a target expanded data set.
Further, the specific manner of step S2 is as follows: and taking the maximum value of the number of targets of each category, and dividing the maximum value by the minimum value of the number of targets of each category, wherein if the result is more than 10, the categories are unbalanced.
Further, the specific manner of step S3 is as follows:
1) calculating the ratio of the number of images of each category to the total number of images [ f 1 ，f 2 ，f 3 ，...，f n ]Defining an expansion threshold t, and screening the category with the proportion smaller than the threshold as a candidate category;
2) screening the imbalance categories in the candidate categories as categories to be expanded, wherein the imbalance categories are determined in the manner described in step S2; the step further screens the category to be expanded in the candidate categories according to the number of samples, and the step can avoid expanding categories with large target number;
3) determining the expansion multiple of each category to be expanded in the following manner:
wherein r is c Denotes the expansion factor of class c, f c Representing the proportion of the image of the c-th class to the total number of the images, wherein t is a threshold value;
4) and expanding the image containing the category to be expanded, wherein the expansion multiple value of the image is as follows:
r i ＝max c∈i (r c )
wherein r is i Representing the current image expansion factor, i representing the image, r c Represents the expansion multiple of the c type;
5) sequentially selecting r from the candidate methods i A transformation method for expanding the current image; the candidate method comprises clockwise 90-degree rotation, clockwise 180-degree rotation, horizontal turning, vertical turning, outward amplification by 10%, inward reduction by 10%, upper left corner clipping, lower right corner clipping, Gaussian noise addition and color transformation.
Compared with the prior art, the invention has the following beneficial effects:
1. the invention provides an improved oversampling method aiming at the defects that simple copying of an image to be subjected to an expansion category easily causes overfitting of a model to a sample and the generalization performance is reduced in the conventional oversampling technology.
2. The invention provides an improved expansion category determining process aiming at the defects that the category to be expanded is unreasonable to select and the category image expansion is carried out under the condition of large target number but intensive distribution in the existing oversampling technology.
Drawings
FIG. 1 is a flowchart of a remote sensing image target detection class imbalance processing method in an embodiment of the present invention.
Fig. 2 is a flow chart of an improved oversampling method.
Detailed Description
The conception, the technical advantages and the technical effects of the present invention will be clearly and completely described in connection with the embodiments, so that the objects, the features and the effects of the present invention can be fully understood. It should be noted that the specific embodiments described herein are only for explaining the present invention, and do not limit the present invention.
A remote sensing image target detection category unbalance processing method based on improved oversampling is provided, which includes firstly counting the number of categories and the number of various targets in a remote sensing data set; then, judging whether the data set has the problem of unbalanced category, if not, ending, and if so, turning to the next step; then, expanding the category images with small target quantity; and finally, obtaining a target expanded data set.
Specifically, as shown in fig. 1, the method comprises the following steps:
s1: counting the number N of categories and the target number [ N ] of each category in the remote sensing data set 1 ，N 2 ，N 3 ，…，N n ]。
S2: judging whether the class imbalance problem exists:
and (4) dividing the maximum value of the number of the targets of each category by the minimum value of the number of the targets of each category, if the result is more than 10, the category is unbalanced, turning to the third step, if the result is less than 10, the category is not unbalanced, and ending the process.
S3: expanding the category images with small target quantity by adopting an improved oversampling method; as shown in fig. 2, the specific manner is as follows:
1) calculating the ratio of the number of the images of each category to the total number of the images [ f 1 ，f 2 ，f 3 ，…，f n ]And artificially defining and expanding a threshold t, and screening the category with the proportion smaller than the threshold as a candidate category.
2) The imbalance category in the candidate categories is screened, and the imbalance category determination method is as described in S2, so as to obtain the category to be expanded.
3) Determining the expansion multiple of each category in the following mode:
wherein r is c Denotes the expansion factor of class c, f c And t is a threshold value set artificially, and represents the proportion of the image of the c-th class to the total number of images.
4) And expanding the image containing the category to be expanded, wherein the expansion multiple value of the image is as follows:
r i ＝max c∈i (r c )
wherein r is i Representing the current image expansion factor, i representing the image, r c Indicating the expansion factor of class c.
5) Expanding the current image, the expanding method sequentially selecting r from the candidate methods i The candidate methods comprise clockwise 90-degree rotation, clockwise 180-degree rotation, horizontal turning, vertical turning, outward amplification by 10%, inward reduction by 10%, upper left corner clipping, lower right corner clipping, Gaussian noise addition and color transformation.
S4: and obtaining a target expanded data set.
In a word, the method can overcome the defect that the prior oversampling technology is easy to cause model overfitting, so that the problem of unbalanced target detection categories of the remote sensing image is solved.
Claims (3)
1. A remote sensing image target detection sample unbalance processing method based on improved oversampling is characterized by comprising the following steps:
s1, counting the number N of categories and the target number [ N ] of each category in the remote sensing data set 1 ，N 2 ，N 3 ，...，N n ]；
S2, judging whether there is the problem of unbalanced classification, if there is unbalanced classification, going to step S3, otherwise, ending the flow;
s3, expanding the category images with small target amount;
s4: and obtaining a target expanded data set.
2. The remote sensing image target detection sample imbalance processing method based on improved oversampling as claimed in claim 1, wherein the specific manner of step S2 is: and taking the maximum value of the number of targets of each category, and dividing the maximum value by the minimum value of the number of targets of each category, wherein if the result is more than 10, the categories are unbalanced.
3. The remote sensing image target detection sample imbalance processing method based on improved oversampling as claimed in claim 2, wherein the specific manner of step S3 is:
1) calculating the ratio of the number of images of each category to the total number of images [ f 1 ，f 2 ，f 3 ，...，f n ]Defining an expansion threshold t, and selecting a category with the screening ratio smaller than the threshold as a candidate category;
2) screening the imbalance categories in the candidate categories as categories to be expanded, wherein the imbalance categories are determined in the manner described in step S2;
3) determining the expansion multiple of each category to be expanded in the following manner:
wherein r is c Denotes the expansion factor of class c, f c Representing the proportion of the c-th image to the total number of images, wherein t is a threshold value;
4) and expanding the image containing the category to be expanded, wherein the expansion multiple value of the image is as follows:
r i ＝max c∈i (r c )
wherein r is i Representing the current image expansion factor, i representing the image, r c Represents the expansion multiple of the c type;
5) selecting r from candidate method in turn i A transformation method for expanding the current image; the candidate method comprises clockwise 90-degree rotation, clockwise 180-degree rotation, horizontal turning, vertical turning, outward amplification by 10%, inward reduction by 10%, upper left corner clipping, lower right corner clipping, Gaussian noise addition and color transformation.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210425413.9A CN114972982A (en) | 2022-04-22 | 2022-04-22 | Remote sensing image target detection sample unbalance processing method based on improved oversampling |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210425413.9A CN114972982A (en) | 2022-04-22 | 2022-04-22 | Remote sensing image target detection sample unbalance processing method based on improved oversampling |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114972982A true CN114972982A (en) | 2022-08-30 |
Family
ID=82979930
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210425413.9A Pending CN114972982A (en) | 2022-04-22 | 2022-04-22 | Remote sensing image target detection sample unbalance processing method based on improved oversampling |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114972982A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117523345A (en) * | 2024-01-08 | 2024-02-06 | 武汉理工大学 | Target detection data balancing method and device |
| CN118262901A (en) * | 2024-04-07 | 2024-06-28 | 中国人民解放军总医院第六医学中心 | Deep learning-based lung cancer type prediction system |
-
2022
- 2022-04-22 CN CN202210425413.9A patent/CN114972982A/en active Pending
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117523345A (en) * | 2024-01-08 | 2024-02-06 | 武汉理工大学 | Target detection data balancing method and device |
| CN117523345B (en) * | 2024-01-08 | 2024-04-23 | 武汉理工大学 | Target detection data balancing method and device |
| CN118262901A (en) * | 2024-04-07 | 2024-06-28 | 中国人民解放军总医院第六医学中心 | Deep learning-based lung cancer type prediction system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN114972982A (en) | Remote sensing image target detection sample unbalance processing method based on improved oversampling | |
| KR102263397B1 (en) | Method for acquiring sample images for inspecting label among auto-labeled images to be used for learning of neural network and sample image acquiring device using the same | |
| CN108765465B (en) | Unsupervised SAR image change detection method | |
| CN109886357B (en) | Feature fusion-based adaptive weight deep learning target classification method | |
| CN113781482B (en) | Method and system for detecting crack defects of mechanical parts in complex environment | |
| CN101923711B (en) | SAR (Synthetic Aperture Radar) image change detection method based on neighborhood similarity and mask enhancement | |
| CN114283287B (en) | Robust field adaptive image learning method based on self-training noise label correction | |
| CN111898685B (en) | Target detection method based on long tail distribution data set | |
| CN112949520B (en) | Aerial photography vehicle detection method and detection system based on multi-scale small samples | |
| CN113469951B (en) | Hub defect detection method based on cascade region convolutional neural network | |
| CN105374047B (en) | SAR image change detection based on improved bilateral filtering with cluster | |
| CN116258707A (en) | PCB surface defect detection method based on improved YOLOv5 algorithm | |
| CN112560614A (en) | Remote sensing image target detection method and system based on candidate frame feature correction | |
| CN110766058A (en) | Battlefield target detection method based on optimized RPN (resilient packet network) | |
| CN112819063B (en) | Image identification method based on improved Focal loss function | |
| CN105184829B (en) | A kind of tight quarters target detection and high-precision method for positioning mass center | |
| CN115457415A (en) | Target detection method and device based on YOLO-X model, electronic equipment and storage medium | |
| CN114743084A (en) | Power transmission line fault detection method based on super-resolution preprocessing and improved YOLOv5x | |
| CN114331950A (en) | SAR image ship detection method based on dense connection sparse activation network | |
| CN114549909A (en) | Pseudo label remote sensing image scene classification method based on self-adaptive threshold | |
| CN115880572A (en) | Forward-looking sonar target identification method based on asynchronous learning factor | |
| CN111368625B (en) | Pedestrian target detection method based on cascade optimization | |
| CN112613462A (en) | Weighted intersection ratio method | |
| CN116597197A (en) | Long-tail target detection method capable of adaptively eliminating negative gradient of classification | |
| CN111325744A (en) | Sample processing method in panel defect detection |
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 |