Weighted adaptive anchor frame cloud particle detection method

Abstract

The invention relates to a weighted adaptive anchor frame cloud particle detection method, which mainly comprises the following steps of: firstly, carrying out different morphological processing on a cloud particle image, extracting multiple types of central points and carrying out first fusion; then, according to the expansion corrosion degree of morphological processing, different weights are given to each fused center point, the weight given to the center point with the larger expansion degree is lower, and the weight given to the center point with the larger corrosion degree is higher; projecting the weighted central points back to the original image, and performing secondary central point fusion and weight updating; and finally, adjusting the size of an anchor frame according to the weight of the central point, generating the anchor frame adaptive to the size of cloud particles, and establishing a deep learning model for detection. According to the method, through adaptive weight and anchor frame size adjustment, the precision of cloud particle target detection is remarkably improved, and the use efficiency of the anchor frame in the detection process is optimized.

Description

Technical Field

[0001] The present invention belongs to the fields of machine learning, image processing and meteorological detection, and in particular relates to detection and positioning of cloud particle image data obtained by airborne equipment through cloud detection based on artificial intelligence methods. Background Art

[0002] Image target detection technology, a key research area in computer vision, has been widely applied in a variety of fields, including military, medicine, agriculture, transportation, and meteorology. In recent years, the widespread development of aircraft cloud penetration experiments has made high-resolution image data collected by cloud particle imagers (CPIs) an essential foundation for cloud microphysics analysis. In particular, detection equipment such as CPI and 2D-S offer extremely high image acquisition speeds and resolutions, providing a rich data resource for analyzing cloud particle characteristics.

[0003] However, due to the scale diversity and uneven distribution of cloud particle targets in actual observations, and the susceptibility of detection equipment to particle fragmentation, pixel loss, and false target interference during imaging, traditional target detection anchor frame generation strategies with fixed sizes and positions cannot effectively match the actual particle size and distribution. This often results in a large number of anchor frames failing to accurately match the target area, leading to redundant detections or missed targets. Furthermore, different morphological processing methods have varying effectiveness in identifying particle regions, and there is a lack of a dynamic, adaptive anchor frame adjustment mechanism tailored to particle morphological differences.

[0004] To address these issues, it is necessary to develop a method for adaptive anchor frame generation based on weighted center points. Initial center points are extracted from multiple morphologically processed images and weighted based on the reliability of each morphologically processed image's features. After fusion, a set of center points with distinct weights is formed. The size of the anchor frames is then dynamically adjusted based on the final weights of the center points. Center points with higher weights correspond to larger anchor frames, while those with lower weights correspond to smaller anchor frames. This weighted adaptive anchor frame generation method effectively matches the varying size characteristics and density distributions of target particles, significantly improving the adaptability and accuracy of cloud particle target detection. Summary of the Invention

[0005] In view of the above-mentioned deficiencies in the prior art, the present invention proposes a weighted adaptive anchor frame cloud particle detection method, which includes the following steps: Step 1: Perform different levels of morphological processing on the original cloud particle image data; Step 2: Generate different types of center points in all morphologically processed cloud particle image data and fuse the center points. Specifically: Step 2.1: Generate different types of center points on all erosion cloud particle image data and expansion cloud particle image data; Step 2.2: Filter out abnormal center points for the center points in each maximum connected area of ​​all erosion cloud particle image data and expansion cloud particle image data; Step 2.3: Fuse the center points retained in each maximum connected area on all erosion cloud particle image data and expansion cloud particle image data respectively; Step 2.4: assign different weights to the center points in the cloud particle image data with different degrees of morphological processing; Step 3: Project the fused center points of all morphologically processed cloud particle image data into the original cloud particle image data, and further fuse the center points. Specifically: Step 3.1: Project the fused center points of all morphologically processed cloud particle image data onto the original cloud particle image data; Step 3.2: Fuse the center points of each largest connected region in the original cloud particle image data; Step 3.3: Repeat step 3.2 until the center point of each maximum connected area in the original cloud particle image data does not meet the fusion condition; Step 4: Use the center point in the original cloud particle image data as the center point of the anchor frame, and combine the weight of the center point to generate different types of anchor frames, specifically: Step 4.1: Take the center point in the original cloud particle image data as the anchor frame center point, and generate different types of anchor frames for each anchor frame center point; Step 4.2: Set the threshold , calculate the number of anchor boxes generated in each original cloud particle image data, when the number of generated anchor boxes is less than the threshold , then by adding a virtual center point in the center of the image to generate a value equal to the threshold The number of anchor boxes; when the number of generated anchor boxes is greater than the threshold , then the most representative anchor box is selected so that the number of generated anchor boxes is equal to the threshold ; Step 5: Build a deep learning model to predict the category probability, center point offset, and anchor box size offset of all generated anchor boxes; Step 6: Filter anchor boxes based on their overlap and category likelihood to obtain detection results. Specifically, assign a category confidence score to each generated anchor box through the classification layer, sort all anchor boxes from high to low according to their category confidence scores, and apply the non-maximum suppression method to filter out and remove redundant anchor boxes based on the intersection-union ratio of the anchor boxes and the category confidence scores to obtain the final detection results.

[0006] Furthermore, the step 1: performing different levels of morphological processing on the original cloud particle image data is specifically as follows: Step 1.1: Binarize the original cloud particle image data; Step 1.2: Invert the original cloud particle image data to obtain inverted cloud particle image data; Step 1.3: Perform M erosion processes on the inverted cloud particle image data to obtain M eroded cloud particle image data; Step 1.4: Perform N expansion processes on the inverted cloud particle image data to obtain N expanded cloud particle image data.

[0007] Furthermore, the step 2.1: generating different types of center points on all corrosion cloud particle image data and expansion cloud particle image data, specifically: generating different types of center points on all corrosion cloud particle image data and expansion cloud particle image data, including geometric center points, minimum circumscribed rectangle center points, maximum connected area center points, contour centroids, and minimum circumscribed circle center points.

[0008] Furthermore, the step 2.2: performing abnormal center point filtering on the center points in each maximum connected area on all corrosion cloud particle image data and expansion cloud particle image data, specifically: calculating the Euclidean distance between all center points in each maximum connected area, and then calculating the average distance between each center point and other center points. Based on the average distance, the outlier method is used to screen out center points with abnormal distances from other center points, and then these center points are removed.

[0009] Furthermore, the step 2.3: fusing the center points retained in each maximum connected region on all erosion cloud particle image data and expansion cloud particle image data, specifically: first calculating the number of pixels in the maximum connected region; then, setting an adaptive threshold based on the number of pixels in the maximum connected region. , , where k is a constant, is the exponential factor, is the number of pixels in the maximum connected area, and then for all the retained center points in each maximum connected area, the Euclidean distance between them is calculated. When the distance between a pair of center points is less than the set threshold , then take the average of the coordinate values ​​of this pair of center points and merge them to get a new center point.

[0010] Furthermore, the step 2.4: assigning different weights to the center points in the cloud particle image data with different degrees of morphological processing, specifically: assigning different weights to the center points of ... The center point in the cloud particle image data is given a weight , , for the expansion times The center point in the cloud particle image data is given a weight , , .

[0011] Furthermore, the step 3.2: fusing the center points of each largest connected region in the original cloud particle image data is as follows: first calculating the number of pixels in the largest connected region; then, setting an adaptive threshold based on the number of pixels in the largest connected region. , ,in, is a constant, is the exponential factor, is the number of pixels in the maximum connected area, and then for all the center points in each maximum connected area, the Euclidean distance between them is calculated. When the distance between a pair of center points is less than the set threshold , then take the average of the coordinate values ​​of this pair of center points, merge them to get a new center point, and assign a new weight to this new center point, which is the sum of the weights of the two center points before the merger.

[0012] Furthermore, the step 4.1: taking the center point in the original cloud particle image data as the anchor frame center point, generating different types of anchor frames for each anchor frame center point, specifically: setting the threshold , set each anchor box center point to generate T anchor boxes, each anchor box uses T1 aspect ratio and T2 scale; when the weight of the anchor box center point is greater than the threshold , then the scales of all generated anchor boxes corresponding to the center point are multiplied by w1, where w1 is the magnification factor; when the weight of the center point of the anchor box is less than or equal to the threshold , then the scales of all generated anchor boxes corresponding to the center point are multiplied by w2, where w2 is the reduction factor; Furthermore, in step 4.2: when the number of generated anchor boxes is greater than the threshold, the most representative anchor boxes are selected so that the number of generated anchor boxes is equal to the threshold, specifically: Step 4.2.1: Group the anchor boxes of the same scale and aspect ratio into a group, calculate the intersection-and-union ratio of the anchor boxes in each group, and sort the intersection-and-union ratio of each group from large to small. Step 4.2.2: For each group of anchor boxes, select the two anchor boxes with the largest intersection-to-union ratio; in each group, randomly delete one of the two anchor boxes with the largest intersection-to-union ratio and keep the other; Step 4.2.3: By repeating step 4.2.2, gradually reduce the redundant anchor boxes until the total number is consistent with the threshold.

[0013] Furthermore, the step 5: establishes a deep learning model to predict the category probability, center point offset and anchor frame size offset of all generated anchor boxes, specifically: first, a partially frozen ResNet-50 network is selected as the backbone network, in which the early layers are frozen and the later layers are unfrozen for feature extraction; then, a two-layer 1x1 convolution transition layer is used to reduce the number of channels of the feature map from 2048 to 512, and the spatial size of the feature map is adjusted through the upsampling mechanism to match the number of generated dynamic anchor boxes. The positioning layer is used to adjust the center point offset and size offset of each anchor box to ensure that the anchor box matches the target area more accurately. The classification layer is used to classify each anchor box and predict the category probability of the anchor box. Finally, the model outputs the category probability, center point offset and anchor frame size offset of each anchor box.

[0014] Compared with traditional cloud particle detection methods, the present invention has the following advantages, thereby solving the corresponding technical problems: 1. The present invention performs different levels of erosion and dilation processing on the original cloud particle image data, ensuring that large-sized cloud particles are continuously highlighted after multiple erosion processes, and small-sized cloud particles are effectively retained through multiple dilation processes, thereby improving the adaptability of target detection to different particle sizes.

[0015] 2. This invention improves the accuracy and reliability of target positioning by extracting multiple types of center points, including geometric center points, minimum circumscribed rectangle center points, maximum connected area center points, contour centroids, and minimum circumscribed circle center points. Furthermore, this invention also designs an outlier point filtering method based on the distance between center points. This method uses an outlier detection mechanism to eliminate anomalous center points, significantly reducing center point positioning errors.

[0016] 3. When performing the first center point fusion on images processed at different levels of morphological processing, each fused center point is assigned a different weight based on the type of morphological processing performed. Images processed with greater erosion have higher weights for the fused center points, while images processed with greater dilation have lower weights for the fused center points. Then, the weighted center points from the first fusion are uniformly projected back into the original cloud particle image data coordinate space. After being projected back into the original image space, a second center point fusion is performed. During the second fusion, center points whose distance is less than a preset threshold are merged, and the weight of the fused center points is equal to the sum of the weights of the fused center points. This weight accumulation mechanism ensures that the center points of large-scale target areas have higher weights, while the center points of small-scale target areas have lower weights.

[0017] 4. We propose a weight-adaptive anchor frame generation mechanism. Based on the center points after secondary fusion, the anchor frame size is adaptively adjusted according to the weight of each center point. For center points with higher weights, the corresponding anchor frame size is increased; for center points with lower weights, the corresponding anchor frame size is reduced. This achieves adaptive matching of the anchor frame size with the actual size of the target particle, significantly reducing the redundancy and omission issues that exist in traditional fixed-size anchor frame generation methods.

[0018] 5. A deep learning model dedicated to cloud particle detection was established. The model can accurately predict the category probability, center point offset, and anchor box size offset of all generated anchor boxes, thereby achieving efficient detection and precise positioning of cloud particle targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 Flowchart of a weighted adaptive anchor box cloud particle detection method. DETAILED DESCRIPTION

[0020] The following will clearly and completely describe the technical solution in the embodiment of the present invention in conjunction with the accompanying drawings in the embodiment of the present invention. The method includes the following steps: Step 1: Perform different levels of morphological processing on the original cloud particle image data; Step 2: Generate different types of center points in all morphologically processed cloud particle image data and fuse the center points. Specifically: Step 2.1: Generate different types of center points on all erosion cloud particle image data and expansion cloud particle image data; Step 2.2: Filter out abnormal center points for the center points in each maximum connected area of ​​all erosion cloud particle image data and expansion cloud particle image data; Step 2.3: Fuse the center points retained in each maximum connected area on all erosion cloud particle image data and expansion cloud particle image data respectively; Step 2.4: assign different weights to the center points in the cloud particle image data with different degrees of morphological processing; Step 3: Project the fused center points of all morphologically processed cloud particle image data into the original cloud particle image data, and further fuse the center points. Specifically: Step 3.1: Project the fused center points of all morphologically processed cloud particle image data onto the original cloud particle image data; Step 3.2: Fuse the center points of each largest connected region in the original cloud particle image data; Step 3.3: Repeat step 3.2 until the center point of each maximum connected area in the original cloud particle image data does not meet the fusion condition; Step 4: Use the center point in the original cloud particle image data as the center point of the anchor frame, and combine the weight of the center point to generate different types of anchor frames, specifically: Step 4.1: Take the center point in the original cloud particle image data as the anchor frame center point, and generate different types of anchor frames for each anchor frame center point; Step 4.2: Set the threshold , calculate the number of anchor boxes generated in each original cloud particle image data, when the number of generated anchor boxes is less than the threshold , then by adding a virtual center point in the center of the image to generate a value equal to the threshold The number of anchor boxes; when the number of generated anchor boxes is greater than the threshold , then the most representative anchor box is selected so that the number of generated anchor boxes is equal to the threshold ; Step 5: Build a deep learning model to predict the category probability, center point offset, and anchor box size offset of all generated anchor boxes; Step 6: Filter anchor boxes based on their overlap and category likelihood to obtain detection results. Specifically, assign a category confidence score to each generated anchor box through the classification layer, sort all anchor boxes from high to low according to their category confidence scores, and apply the non-maximum suppression method to filter out and remove redundant anchor boxes based on the intersection-union ratio of the anchor boxes and the category confidence scores to obtain the final detection results.

[0021] Furthermore, the step 1: performing different levels of morphological processing on the original cloud particle image data is specifically as follows: Step 1.1: Binarize the original cloud particle image data; Step 1.2: Invert the original cloud particle image data to obtain inverted cloud particle image data; Step 1.3: Perform M erosion processes on the inverted cloud particle image data to obtain M eroded cloud particle image data; Step 1.4: Perform N expansion processes on the inverted cloud particle image data to obtain N expanded cloud particle image data.

[0022] Furthermore, the step 2.1: generating different types of center points on all corrosion cloud particle image data and expansion cloud particle image data, specifically: generating different types of center points on all corrosion cloud particle image data and expansion cloud particle image data, including geometric center points, minimum circumscribed rectangle center points, maximum connected area center points, contour centroids, and minimum circumscribed circle center points.

[0023] Furthermore, the step 2.2: performing abnormal center point filtering on the center points in each maximum connected area on all corrosion cloud particle image data and expansion cloud particle image data, specifically: calculating the Euclidean distance between all center points in each maximum connected area, and then calculating the average distance between each center point and other center points. Based on the average distance, the outlier method is used to screen out center points with abnormal distances from other center points, and then these center points are removed.

[0024] Furthermore, the step 2.3: fusing the center points retained in each maximum connected region on all erosion cloud particle image data and expansion cloud particle image data, specifically: first calculating the number of pixels in the maximum connected region; then, setting an adaptive threshold based on the number of pixels in the maximum connected region. , , where k is a constant, is the exponential factor, is the number of pixels in the maximum connected area, and then for all the retained center points in each maximum connected area, the Euclidean distance between them is calculated. When the distance between a pair of center points is less than the set threshold , then take the average of the coordinate values ​​of this pair of center points and merge them to get a new center point.

[0025] Furthermore, the step 2.4: assigning different weights to the center points in the cloud particle image data with different degrees of morphological processing, specifically: assigning different weights to the center points of ... The center point in the cloud particle image data is given a weight , , for the expansion times The center point in the cloud particle image data is given a weight , , .

[0026] Furthermore, the step 3.2: fusing the center points of each largest connected region in the original cloud particle image data is as follows: first calculating the number of pixels in the largest connected region; then, setting an adaptive threshold based on the number of pixels in the largest connected region. , ,in, is a constant, is the exponential factor, is the number of pixels in the maximum connected area, and then for all the center points in each maximum connected area, the Euclidean distance between them is calculated. When the distance between a pair of center points is less than the set threshold , then take the average of the coordinate values ​​of this pair of center points, merge them to get a new center point, and assign a new weight to this new center point, which is the sum of the weights of the two center points before the merger.

[0027] Furthermore, the step 4.1: taking the center point in the original cloud particle image data as the anchor frame center point, generating different types of anchor frames for each anchor frame center point, specifically: setting the threshold , set each anchor box center point to generate T anchor boxes, each anchor box uses T1 aspect ratio and T2 scale; when the weight of the anchor box center point is greater than the threshold , then the scales of all generated anchor boxes corresponding to the center point are multiplied by w1, where w1 is the magnification factor; when the weight of the center point of the anchor box is less than or equal to the threshold , then the scales of all generated anchor boxes corresponding to the center point are multiplied by w2, where w2 is the reduction factor; Furthermore, in step 4.2: when the number of generated anchor boxes is greater than the threshold, the most representative anchor boxes are selected so that the number of generated anchor boxes is equal to the threshold, specifically: Step 4.2.1: Group the anchor boxes of the same scale and aspect ratio into a group, calculate the intersection-and-union ratio of the anchor boxes in each group, and sort the intersection-and-union ratio of each group from large to small. Step 4.2.2: For each group of anchor boxes, select the two anchor boxes with the largest intersection-to-union ratio; in each group, randomly delete one of the two anchor boxes with the largest intersection-to-union ratio and keep the other; Step 4.2.3: By repeating step 4.2.2, gradually reduce the redundant anchor boxes until the total number is consistent with the threshold.

[0028] Furthermore, the step 5: establishes a deep learning model to predict the category probability, center point offset and anchor frame size offset of all generated anchor boxes, specifically: first, a partially frozen ResNet-50 network is selected as the backbone network, in which the early layers are frozen and the later layers are unfrozen for feature extraction; then, a two-layer 1x1 convolution transition layer is used to reduce the number of channels of the feature map from 2048 to 512, and the spatial size of the feature map is adjusted through the upsampling mechanism to match the number of generated dynamic anchor boxes. The positioning layer is used to adjust the center point offset and size offset of each anchor box to ensure that the anchor box matches the target area more accurately. The classification layer is used to classify each anchor box and predict the category probability of the anchor box. Finally, the model outputs the category probability, center point offset and anchor frame size offset of each anchor box.

[0029] Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, other embodiments obtained by ordinary technicians in this field without making creative work are all within the scope of protection of the present invention.

Claims

1. A weighted adaptive anchor box cloud particle detection method, comprising the following steps: Step 1: Perform different levels of morphological processing on the original cloud particle image data; Step 2: Generate different types of center points in all morphologically processed cloud particle image data and fuse the center points. Specifically: Step 2.1: Generate different types of center points on all erosion cloud particle image data and expansion cloud particle image data; Step 2.2: Filter out abnormal center points for the center points in each maximum connected area of ​​all erosion cloud particle image data and expansion cloud particle image data; Step 2.3: Fuse the center points retained in each maximum connected area on all erosion cloud particle image data and expansion cloud particle image data respectively; Step 2.4: assign different weights to the center points in the cloud particle image data with different degrees of morphological processing; Step 3: Project the fused center points of all morphologically processed cloud particle image data into the original cloud particle image data, and further fuse the center points. Specifically: Step 3.1: Project the fused center points of all morphologically processed cloud particle image data onto the original cloud particle image data; Step 3.2: Fuse the center points of each largest connected region in the original cloud particle image data; Step 3.3: Repeat step 3.2 until the center point of each maximum connected area in the original cloud particle image data does not meet the fusion condition; Step 4: Use the center point in the original cloud particle image data as the center point of the anchor frame, and combine the weight of the center point to generate different types of anchor frames, specifically: Step 4.1: Take the center point in the original cloud particle image data as the anchor frame center point, and generate different types of anchor frames for each anchor frame center point; Step 4.2: Set the threshold , calculate the number of anchor boxes generated in each original cloud particle image data, when the number of generated anchor boxes is less than the threshold , then by adding a virtual center point in the center of the image to generate a value equal to the threshold The number of anchor boxes; When the number of generated anchor boxes is greater than the threshold , then the most representative anchor box is selected so that the number of generated anchor boxes is equal to the threshold ; Step 5: Build a deep learning model to predict the category probability, center point offset, and anchor box size offset of all generated anchor boxes; Step 6: Filter anchor boxes based on their overlap and category likelihood to obtain detection results. Specifically, assign a category confidence score to each generated anchor box through the classification layer, sort all anchor boxes from high to low according to their category confidence scores, and apply the non-maximum suppression method to filter out and remove redundant anchor boxes based on the intersection-union ratio of the anchor boxes and the category confidence scores to obtain the final detection results.

2. A weighted adaptive anchor frame cloud particle detection method as claimed in claim 1, characterized in that Step 1: performing different levels of morphological processing on the original cloud particle image data, specifically: Step 1.1: Binarize the original cloud particle image data; Step 1.2: Invert the original cloud particle image data to obtain inverted cloud particle image data; Step 1.3: Perform M erosion processes on the inverted cloud particle image data to obtain M eroded cloud particle image data; Step 1.4: Perform N expansion processes on the inverted cloud particle image data to obtain N expanded cloud particle image data.

3. A weighted adaptive anchor frame cloud particle detection method as claimed in claim 1, characterized in that The step 2.1: generating different types of center points on all corrosion cloud particle image data and expansion cloud particle image data, specifically: generating different types of center points on all corrosion cloud particle image data and expansion cloud particle image data, including geometric center points, minimum circumscribed rectangle center points, maximum connected area center points, contour centroids, and minimum circumscribed circle center points.

4. A weighted adaptive anchor frame cloud particle detection method as claimed in claim 1, characterized in that The step 2.2: performing abnormal center point filtering on the center points in each maximum connected area of ​​all corrosion cloud particle image data and expansion cloud particle image data, specifically: calculating the Euclidean distance between all center points in each maximum connected area, and then calculating the average distance between each center point and other center points. Based on the average distance, an outlier method is used to filter out center points with abnormal distances from other center points, and then these center points are removed.

5. A weighted adaptive anchor frame cloud particle detection method as claimed in claim 1, characterized in that Step 2.3: Fusing the center points retained in each maximum connected region on all erosion cloud particle image data and expansion cloud particle image data, specifically: first calculating the number of pixels in the maximum connected region; then, setting an adaptive threshold based on the number of pixels in the maximum connected region. , , where k is a constant, is the exponential factor, is the number of pixels in the maximum connected area, and then for all the retained center points in each maximum connected area, the Euclidean distance between them is calculated. When the distance between a pair of center points is less than the set threshold , then take the average of the coordinate values ​​of this pair of center points and merge them to get a new center point.

6. A weighted adaptive anchor frame cloud particle detection method as claimed in claim 1, characterized in that Step 2.4: assign different weights to the center points in the cloud particle image data with different degrees of morphological processing, specifically: assign different weights to the center points with the number of corrosion times The center point in the cloud particle image data is given a weight , , for the expansion times The center point in the cloud particle image data is given a weight , , .

7. A weighted adaptive anchor frame cloud particle detection method as claimed in claim 1, characterized in that Step 3.2: Fusing the center points of each largest connected region in the original cloud particle image data, specifically: first calculating the number of pixels in the largest connected region; then, setting an adaptive threshold based on the number of pixels in the largest connected region. , ,in, is a constant, is the exponential factor, is the number of pixels in the maximum connected area, and then for all the center points in each maximum connected area, the Euclidean distance between them is calculated. When the distance between a pair of center points is less than the set threshold , then take the average of the coordinate values ​​of this pair of center points, merge them to get a new center point, and assign a new weight to this new center point, which is the sum of the weights of the two center points before the merger.

8. A weighted adaptive anchor frame cloud particle detection method as claimed in claim 1, characterized in that Step 4.1: Using the center point in the original cloud particle image data as the anchor frame center point, generate different types of anchor frames for each anchor frame center point, specifically: setting the threshold , set each anchor box center point to generate T anchor boxes, each anchor box uses T1 aspect ratio and T2 scale; when the weight of the anchor box center point is greater than the threshold , then the scales of all generated anchor boxes corresponding to the center point are multiplied by w1, where w1 is the magnification factor; when the weight of the center point of the anchor box is less than or equal to the threshold , then the scales of all generated anchor boxes corresponding to the center point are multiplied by w2, where w2 is the reduction factor.

9. A weighted adaptive anchor frame cloud particle detection method as claimed in claim 1, characterized in that Step 4.2: When the number of generated anchor boxes is greater than the threshold, the most representative anchor boxes are selected so that the number of generated anchor boxes is equal to the threshold. Specifically: Step 4.2.1: Group the anchor boxes of the same scale and aspect ratio into a group, calculate the intersection-and-union ratio of the anchor boxes in each group, and sort the intersection-and-union ratio of each group from large to small. Step 4.2.2: For each group of anchor boxes, select the two anchor boxes with the largest intersection-to-union ratio; in each group, randomly delete one of the two anchor boxes with the largest intersection-to-union ratio and keep the other; Step 4.2.3: By repeating step 4.2.2, gradually reduce the redundant anchor boxes until the total number is consistent with the threshold.

10. A weighted adaptive anchor frame cloud particle detection method according to claim 1, characterized in that The step 5: establishing a deep learning model to predict the category probability, center point offset and anchor frame size offset of all generated anchor boxes, specifically: first, selecting a partially frozen ResNet-50 network as the backbone network, in which the early layers are frozen and the later layers are unfrozen for feature extraction; then using a two-layer 1x1 convolution transition layer to reduce the number of channels of the feature map from 2048 to 512, adjusting the spatial size of the feature map through an upsampling mechanism to match the number of generated dynamic anchor boxes, using a positioning layer to adjust the center point offset and size offset of each anchor box to ensure that the anchor box matches the target area more accurately, using a classification layer to classify each anchor box and predict the category probability of the anchor box, and finally the model outputs the category probability, center point offset and anchor frame size offset of each anchor box.

Images