CN-122023907-A - Target detection and recognition method based on feature clustering subdivision class

CN122023907ACN 122023907 ACN122023907 ACN 122023907ACN-122023907-A

Abstract

The invention provides a target detection and recognition method based on feature clustering subdivision categories, and relates to the technical fields of image processing, target detection and machine learning. The method comprises the steps of obtaining an original sample image, cutting out a label area in the original sample image, calculating a statistical feature vector of a slice image, carrying out feature clustering on the slice image of each original category, giving a new category label to the slice image according to a subdivision category, training a target detection model by taking the slice image with the new category label as a sample set, carrying out target detection on the image to be detected by using the trained target detection model, and restoring the target category to the corresponding original category according to the corresponding relation between the subdivision category and the original category to obtain the final category of the target. The method and the device can effectively capture the significant characteristic difference between samples, reduce the difference in categories, increase the difference between the categories and improve the detection performance of the target detection model.

Inventors

LI YONGQUAN
ZHU CHANGREN
CHEN JIANGTAO
Zhao Huipan
ZHAO XIANGYANG
YANG YINMING
ZHANG XU
YANG XIAOLIANG
ZHANG GANG

Assignees

中国电子科技集团公司第五十四研究所

Dates

Publication Date: 20260512
Application Date: 20260130

Claims (4)

1. The target detection and identification method based on the feature clustering subdivision class is characterized by comprising the following steps of: step 1, an original sample image is obtained, wherein the image is provided with a label area, and the labeling category of the label area is the original category; Step 2, cutting out a label area in an original sample image, marking the label area as a slice image, and calculating various statistic characteristics of each slice image to form a statistic characteristic vector of the slice image; Step 3, based on the statistical feature vector, carrying out feature clustering on the slice images of each original category to realize subdivision of the original category; Step 4, taking the slice image with the new class label as a sample set, and training the target detection model to obtain a trained target detection model; Step 5, performing target detection on the image to be detected by using the trained target detection model, and identifying the target category in the image; and 6, restoring the target category identified in the step 5 into a corresponding original category according to the corresponding relation between the subdivision category and the original category, obtaining the final category of the target, and completing target detection and identification.
2. The method according to claim 1, wherein in step 2, the plurality of statistic features include mean, median, variance, skewness, kurtosis, covariance feature vector, and histogram distribution feature vector of the slice image.
3. The target detection and recognition method based on feature clustering subdivision category as claimed in claim 1, wherein in step 3, feature clustering is performed by adopting MEAN SHIFT method, specifically, the method comprises the following steps: Step 301, taking the statistical feature vector of each slice image in the same original category as a sample point, and forming a set from the sample points in the same original category , Is the number of sample points; Step 302, constructing a sample point sequence Initializing ; Step 303, calculating bandwidth parameters : Wherein, the Representation of Each point in (a) From small to large row of distance values of the q-th bit, , Representing a downward rounding; The distance is calculated by the following steps: , wherein, 、 Is the two points in the X, Is that And (3) with The distance between the two plates is set to be equal, Representing a 2-norm; Step 304, calculate a sample point sequence Updated values for respective points in (a) : Wherein, the Representation of Is used in the neighborhood of (a), , Representing a neighborhood The number of all points in (a); step 305, verifying whether the following conditions are met: Wherein max represents taking the maximum value; and for a sequence of sample points Is updated at each point in the list: ; If the condition is satisfied, executing step 306 after updating, otherwise, returning to step 304 after updating; step 306, traversing Points in (a) If (if) The clustering condition is satisfied: Will be Is a neighborhood of (a) As a cluster Will be From the point in (a) Continuing traversing the rest points in the X until all the rest points do not meet the clustering condition, wherein the set of the rest points is S and the set of the clustering points is M; step 307, for each If there is So that Will then Falls into clusters ; Finally, all the obtained clusters are the subdivision categories of the corresponding original categories.
4. The method for detecting and identifying the target based on the feature cluster subdivision category according to claim 1, wherein in step 4, the target detection model is YOLOv s model, and the training framework provided by Ultralytics is adopted to train the target detection model.

Description

Target detection and recognition method based on feature clustering subdivision class Technical Field The invention relates to the technical fields of image processing, target detection and machine learning, in particular to a target detection and identification method based on feature clustering subdivision categories. Background In the target detection task, the accuracy and fineness of the labels directly influence the performance of the model. However, the existing target detection data set generally adopts a rough label classification mode, so that samples in the same label category have large differences, and the distinction degree between different label categories is insufficient. This phenomenon may significantly reduce the recognition accuracy and generalization ability of the object detection model. Taking the unmanned aerial vehicle image as an example, the vehicle photographed in the daytime usually has higher resolution and rich color information, and at night or in shadow, the image may become blurred or noisy due to insufficient light, so that it is difficult for the object detection model to accurately identify the type of the vehicle. In addition, existing label classification methods are not fine enough and in practice often fall different types of vehicles (e.g., buses, trucks, trailers, cars, vans, SUVs, etc.) into the same category (e.g., "large vehicles" or "small vehicles") without consideration of their respective characteristics. This rough way of label classification makes it difficult for the model to meet the needs of diversification in practical applications, especially in complex scenarios, such as crowded city streets or different weather conditions, where the performance of the model is further degraded. Furthermore, the prior art is mainly dependent on manual labeling or simple rule partitioning methods for tag subdivision categories. These methods are not only time consuming and labor intensive, but are also not necessarily reliable. Therefore, how to realize label subdivision category by an automatic method, reduce the variability in the same category, and enhance the degree of distinction between different categories becomes a problem to be solved in the current target detection field. Disclosure of Invention Aiming at the defects of the prior art, the invention provides a target detection and identification method based on feature clustering subdivision categories. The method can effectively capture the significant characteristic difference between samples, reduce the difference in categories, increase the difference between the categories, tamp the basis for the detection performance of the subsequent target detection model, and improve the detection performance of the target detection model. The invention adopts the technical scheme that: A target detection and identification method based on feature clustering subdivision category comprises the following steps: step 1, an original sample image is obtained, wherein the image is provided with a label area, and the labeling category of the label area is the original category; Step 2, cutting out a label area in an original sample image, marking the label area as a slice image, and calculating various statistic characteristics of each slice image to form a statistic characteristic vector of the slice image; Step 3, based on the statistical feature vector, carrying out feature clustering on the slice images of each original category to realize subdivision of the original category; Step 4, taking the slice image with the new class label as a sample set, and training the target detection model to obtain a trained target detection model; Step 5, performing target detection on the image to be detected by using the trained target detection model, and identifying the target category in the image; and 6, restoring the target category identified in the step 5 into a corresponding original category according to the corresponding relation between the subdivision category and the original category, obtaining the final category of the target, and completing target detection and identification. Further, in step 2, the plurality of statistic features include mean, median, variance, skewness, kurtosis, covariance feature vector, histogram distribution feature vector of the slice image. Further, in step 3, a MEAN SHIFT method is adopted to perform feature clustering, and the specific method is as follows: Step 301, taking the statistical feature vector of each slice image in the same original category as a sample point, and forming a set from the sample points in the same original category ,Is the number of sample points; Step 302, constructing a sample point sequence Initializing; Step 303, calculating bandwidth parameters: Wherein, the Representation ofEach point in (a)From small to large row of distance values of the q-th bit,,Representing a downward rounding; The distance is calculated by the following steps: , wherein, 、Is the two p