Target detection and identification method based on feature clustering classification subdivision

Abstract

The invention provides a target detection and recognition method based on feature clustering classification subdivision, and relates to the technical field 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, and calculating a statistical feature vector of a slice image; feature clustering is carried out on the slice images of each original category, and new category labels are given to the slice images according to subdivided categories; taking the slice images with the new category labels as a sample set, and training a target detection model; performing target detection on a to-be-detected image by using the trained target detection model; and restoring the target category to the corresponding original category according to the corresponding relationship between the subdivided category and the original category, and obtaining the final category of the target. The method can effectively capture significant feature differences between samples, reduce differences in classes, increase differences between classes, and improve the detection performance of a target detection model.

Description

Technical Field

[0001] This invention relates to the fields of image processing, object detection, and machine learning, and specifically to an object detection and recognition method based on feature clustering for subdividing categories. Background Technology

[0002] In object detection tasks, the accuracy and precision of labels directly impact model performance. However, existing object detection datasets typically employ coarse label classification methods, leading to significant differences among samples within the same label category and insufficient discriminative power between different label categories. This phenomenon significantly reduces the recognition accuracy and generalization ability of object detection models. For example, drone images captured during the day usually have high resolution and rich color information, while at night or in shadows, due to insufficient light, images may become blurry or noisy, making it difficult for object detection models to accurately identify vehicle types. Furthermore, existing label classification methods are not precise enough; in practice, different types of vehicles (such as buses, trucks, trailers, cars, vans, SUVs, etc.) are often grouped into the same category (e.g., "large vehicles" or "small vehicles") without considering their individual characteristics. This coarse label classification makes it difficult for models to meet diverse needs in real-world applications, especially in complex scenarios such as congested city streets or varying weather conditions, where model performance further deteriorates.

[0003] Furthermore, current technologies for subdividing labels mainly rely on manual annotation or simple rule-based classification methods. These methods are not only time-consuming and labor-intensive, but also not always reliable. Therefore, how to automate the subdivision of labels, reduce the differences within the same category, and enhance the distinguishability between different categories has become an urgent problem to be solved in the field of object detection. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a target detection and recognition method based on feature clustering for further category segmentation. This invention effectively captures significant feature differences between samples, reduces intra-category differences, and increases inter-category differences, thus laying a solid foundation for the detection performance of subsequent target detection models and improving their overall detection performance.

[0005] The technical solution adopted in this invention is as follows:

[0006] A target detection and recognition method based on feature clustering for further category segmentation includes the following steps:

[0007] Step 1: Obtain the original sample image, which contains labeled regions. The labeled regions are classified as the original categories.

[0008] Step 2: Crop the label region from the original sample image and denote it as a slice image. Calculate the various statistical features of each slice image to form the statistical feature vector of the slice image.

[0009] Step 3: Based on statistical feature vectors, perform feature clustering on the slice images of each original category to achieve subdivision of the original category; assign new category labels to the slice images according to the subdivision categories;

[0010] Step 4: Use the sliced ​​images with new category labels as a sample set to train the object detection model and obtain the trained object detection model.

[0011] Step 5: Use the trained object detection model to perform object detection on the image to be detected and identify the object category in the image;

[0012] Step 6: Based on the correspondence between the subcategories and the original categories, restore the target category identified in Step 5 to the corresponding original category to obtain the final category of the target and complete the target detection and recognition.

[0013] Furthermore, in step 2, the various statistical features include the mean, median, variance, skewness, kurtosis, covariance feature vector, and histogram distribution feature vector of the slice image.

[0014] Furthermore, in step 3, the Mean Shift method is used for feature clustering, specifically as follows:

[0015] Step 301: Using the statistical feature vector of each slice image in the same original category as a sample point, form a set of sample points under the same original category. , The number of sample points;

[0016] Step 302, construct the sample point sequence ,initialization ;

[0017] Step 303, calculate bandwidth parameters :

[0018]

[0019] in, express various points in the middle and The distance value that is qth in order of magnitude from smallest to largest. , Indicates rounding down;

[0020] The distance is calculated as follows: ,in, , Let X be two points. for and The distance between them Represents the 2-norm;

[0021] Step 304, calculate the sample point sequence Update values ​​of each point :

[0022]

[0023] in, express The neighborhood, , Representing the neighborhood The number of all points;

[0024] Step 305, verify whether the following conditions are met:

[0025]

[0026] Where max represents the maximum value;

[0027] And the sample point sequence Update each point in the data: ;

[0028] If the conditions are met, proceed to step 306 after the update; otherwise, return to step 304 after the update.

[0029] Step 306, Traverse Points in ,like Clustering conditions are met:

[0030]

[0031] Then neighborhood As a cluster ,Will The point in Delete the remaining points in X and continue to traverse the remaining points in X until none of the remaining points satisfy the clustering condition. The set of remaining points is S and the set of clustering points is M.

[0032] Step 307, for each If it exists , making Then Clustering ;

[0033] Ultimately, all the resulting clusters are the subcategories of the corresponding original categories.

[0034] Furthermore, in step 4, the target detection model is a YOLOv11s model, and the target detection model is trained using the training framework provided by Ultralytics.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] 1. This invention achieves automated segmentation of target detection label categories by combining slice statistical feature vector extraction and cluster analysis, significantly reducing the differences within the same label category and enhancing the distinguishability between different categories.

[0037] 2. This invention proposes a method for feature clustering and subdividing categories for each type of label, which can more accurately capture the feature differences between different categories, thereby improving the performance of the object detection model.

[0038] 3. This invention replaces traditional manual labeling or simple rule-based classification methods with automated processing, which greatly reduces the workload and cost of subdividing labels and improves data processing efficiency. Attached Figure Description

[0039] Figure 1 This is a schematic diagram illustrating the principle of the method in an embodiment of the present invention. Detailed Implementation

[0040] The present invention will now be described in further detail with reference to the accompanying drawings.

[0041] A target detection and recognition method based on feature clustering for further subdivision of categories, such as Figure 1 As shown, it includes the following steps:

[0042] Step 1: Obtain raw sample images from open source or manually collected sources. These images contain labeled regions, and the labeled categories of these regions will be referred to as the raw categories.

[0043] Step 2: Crop the label region from the original sample image and record it as a slice image. The number of slice images shall not be less than 100. Calculate the various statistical features of each slice image to form the statistical feature vector of the slice image.

[0044] Assuming the sliced ​​image size is (C, H, W), where C represents the number of channels, H represents the height, and W represents the width; various statistical features include:

[0045] 1) Mean of sliced ​​images :

[0046]

[0047] in, This represents the pixel value at channel k, pixel height i, and pixel width j, and so on.

[0048] 2) Median of sliced ​​images :

[0049]

[0050] in, This represents a sorted array of pixel values. This means taking the first element of the array. The pixel value of the bit. Indicates the number Round down to the nearest integer. express The length.

[0051] 3) Variance of sliced ​​images :

[0052]

[0053] 4) Skewness of sliced ​​images :

[0054]

[0055] 5) Kurtosis of the slice image :

[0056]

[0057] 6) Covariance eigenvectors of sliced ​​images :

[0058]

[0059]

[0060]

[0061] in, Indicates the image number 1 The average of each channel, Represents image channels The sample covariance between them. For a single... image, In the calculation formula The value is usually 3.

[0062] 7) Histogram distribution feature vector of sliced ​​image :

[0063]

[0064]

[0065]

[0066]

[0067]

[0068] in, , Histogram representing hyperparameter settings Quantity, range of values ​​is .

[0069] Finally, statistical feature vectors for:

[0070]

[0071] in, .

[0072] This step involves various digital image processing techniques. By performing calculations on these label slice images, preliminary prior knowledge of the label slice images is obtained.

[0073] Step 3: Based on statistical feature vectors, perform feature clustering on the slice images of each original category to achieve subdivision of the original category; assign new category labels to the slice images according to the subdivision categories;

[0074] This step uses the Mean Shift algorithm for clustering. The Mean Shift algorithm is an unsupervised machine learning algorithm primarily used for data clustering. It iteratively adjusts the positions of data points, causing each data point to move towards the mean of its region, eventually converging to the center of the high-density region, which is the cluster center. The clustering method is as follows:

[0075] Step 301: Using the statistical feature vector of each slice image in the same original category as a sample point, form a set of sample points under the same original category. , The number of sample points;

[0076] Step 302, construct the sample point sequence ,initialization ;

[0077] Step 303, calculate bandwidth parameters :

[0078]

[0079] in, express various points in the middle and The distance value that is qth in order of magnitude from smallest to largest. ;

[0080] The distance is calculated as follows: ,in, , Let X be two points. for and The distance between them Represents the 2-norm;

[0081] Step 304, calculate the sample point sequence Update values ​​of each point :

[0082]

[0083] in, express The neighborhood of a point is the point in the neighborhood of a point. Distance not exceeding All points, i.e.:

[0084]

[0085] Representing the neighborhood The number of all points;

[0086] Step 305, verify whether the following conditions are met:

[0087]

[0088] Where max represents the maximum value;

[0089] And the sample point sequence Update each point in the data: ;

[0090] If the conditions are met, proceed to step 306 after the update; otherwise, return to step 304 after the update.

[0091] Step 306, Traverse Points in ,like Clustering conditions are met:

[0092]

[0093] Then neighborhood As a cluster ,Will The point in Delete the remaining points in X and continue to traverse the remaining points in X until none of the remaining points satisfy the clustering condition. The set of remaining points is S and the set of clustering points is M.

[0094] Step 307, for each If it exists , making Then Clustering ;

[0095] Ultimately, all the resulting clusters are the subcategories of the corresponding original categories.

[0096] This step achieves more accurate label segmentation by performing feature clustering to further subdivide the categories for each label.

[0097] Step 4: Using the sliced ​​images with new category labels as a sample set, train the YOLOv11s object detection model using the Ultralytics training framework; further, the trained object detection model can be tested to obtain the trained object detection model.

[0098] This step uses YOLOv11s as the base detection model. This model is a publicly available official version, and its structure and parameter configuration follow the standard YOLOv11s architecture. YOLOv11s is a lightweight object detection model with fast inference speed and good detection accuracy. While maintaining high accuracy, it significantly reduces the number of parameters and computational complexity, making it suitable for real-time object detection scenarios.

[0099] The Ultralytics training framework is a publicly available official version that uses a combination of cross-entropy loss function and mean squared error loss function to improve the model's recognition accuracy for different categories. It also introduces data augmentation techniques (such as random cropping, rotation, and color jitter) during the training phase to increase sample diversity and improve the model's generalization ability.

[0100] This step allows you to evaluate the trained YOLOv11s model using a separate test set. This test set contains new image data with the same structure as the training set but which was not used in the training.

[0101] Step 5: Input the image to be detected into the trained object detection model and obtain the prediction result of the category to which each object box belongs. At this time, the category label is the sub-category label.

[0102] Step 6: Based on the correspondence between the subcategories and the original categories, restore the subcategories identified in Step 5 to their corresponding original categories, which are the final categories of the target.

[0103] At this point, the target detection and recognition are complete.

[0104] This invention achieves detailed classification of target detection labels through two main steps: slice statistical feature vector extraction and statistical feature vector clustering. This addresses the problems of imprecise label classification and limited model performance in existing target detection technologies. This method not only reduces the workload of manual annotation but also improves data processing efficiency, providing an effective solution for complex scenarios in practical applications.

[0105] This invention can effectively capture significant feature differences between samples, reduce intra-class differences, and increase inter-class differences, laying a solid foundation for the detection performance of subsequent target detection models and improving the detection performance of target detection models.

Claims

1. A target detection and recognition method based on feature clustering for subdividing categories, characterized in that, Includes the following steps: Step 1: Obtain the original sample image, which contains labeled regions. The labeled regions are classified as the original categories. Step 2: Crop the label region from the original sample image and denote it as a slice image. Calculate the various statistical features of each slice image to form the statistical feature vector of the slice image. Step 3: Based on statistical feature vectors, perform feature clustering on the slice images of each original category to achieve further subdivision of the original categories; Assign new category labels to the sliced ​​images based on the subcategories; Step 4: Use the sliced ​​images with new category labels as a sample set to train the object detection model and obtain the trained object detection model. Step 5: Use the trained object detection model to perform object detection on the image to be detected and identify the object category in the image; Step 6: Based on the correspondence between the subcategories and the original categories, restore the target category identified in Step 5 to the corresponding original category to obtain the final category of the target and complete the target detection and recognition.

2. The target detection and recognition method based on feature clustering for subdividing categories according to claim 1, characterized in that, In step 2, the various statistical features include the 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 for subdividing categories according to claim 1, characterized in that, In step 3, the Mean Shift method is used for feature clustering, specifically as follows: Step 301: Using the statistical feature vector of each slice image in the same original category as a sample point, form a set of sample points under the same original category. , The number of sample points; Step 302, construct the sample point sequence ,initialization ; Step 303, calculate bandwidth parameters : in, express various points in the middle and The distance value that is qth in order of magnitude from smallest to largest. , Indicates rounding down; The distance is calculated as follows: ,in, , Let X be two points. for and The distance between them Represents the 2-norm; Step 304, calculate the sample point sequence Update values ​​of each point : in, express The neighborhood, , Representing the neighborhood The number of all points; Step 305, verify whether the following conditions are met: Where max represents the maximum value; And the sample point sequence Update each point in the data: ; If the conditions are met, proceed to step 306 after the update; otherwise, return to step 304 after the update. Step 306, Traverse Points in ,like Clustering conditions are met: Then neighborhood As a cluster ,Will The point in Delete the remaining points in X and continue to traverse the remaining points in X until none of the remaining points satisfy the clustering condition. The set of remaining points is S and the set of clustering points is M. Step 307, for each If it exists , making Then Clustering ; Ultimately, all the resulting clusters are the subcategories of the corresponding original categories.

4. The target detection and recognition method based on feature clustering for subdividing categories according to claim 1, characterized in that, In step 4, the target detection model is the YOLOv11s model, and the target detection model is trained using the training framework provided by Ultralytics.

Images