Cascade center point fusion cloud particle detection method
Through the cascade center point fusion method and deep learning model, the problem of inaccurate detection of airborne cloud particle detection equipment in complex environments was solved, and high-precision and stable cloud particle area detection and classification were achieved.
Patent Information
- Application Number
- CN202510738210.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-04
- Publication Date
- 2025-09-16
AI Technical Summary
With existing technology, when an aircraft passes through clouds, the onboard cloud particle detection equipment is affected by instrument vibration, particle motion and environmental interference, resulting in particle breakage, false targets and obstructions, making it difficult to accurately detect and classify cloud particle areas. In addition, a single scale or morphological method is not sufficient to fully capture particle characteristics.
The cascade center point fusion method is adopted to generate the initial cloud particle center points through morphological processing of different scales and shapes. Multi-level fusion is performed and combined with the deep learning model to extract the spatial distribution information of cloud particles, generate anchor frames and perform screening to finally obtain the detection results.
It improves the stability and accuracy of cloud particle area detection, enhances the accuracy of particle classification and identification, adapts to scale differences in complex scenes, reduces the amount of calculation and improves the robustness of detection results.
Smart Images
Figure CN120655896A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of cloud particle target detection in meteorological detection and deep learning, and in particular relates to a cascade center point fusion cloud particle detection method. Background Art
[0002] As aircraft cloud penetration experiments become an increasingly important research tool in cloud physics observation, high-resolution particle image data acquired by airborne cloud particle detection equipment is becoming a crucial foundational data source for cloud microphysics research. Aircraft-mounted imaging equipment, such as CPI and 2D-S probes, can directly capture intuitive images of cloud particles, providing intuitive data support for studying particle physical properties, particle spectral distribution characteristics, liquid water content, ice crystal size, and concentration distribution.
[0003] However, due to the high-speed motion of airborne detection equipment, which is affected by instrument vibration, rapid particle motion, and environmental interference, actual observations are subject to interference such as particle fragmentation, false targets, particle occlusion, and internal voids, which seriously affect the precise detection, positioning, and classification of cloud particle regions. Traditional detection methods such as template matching or morphological processing often perform poorly in dealing with these interferences, and are prone to detecting split cloud particles as multiple particles. In addition, due to the complexity and diversity of cloud particles, it is difficult to fully and accurately capture the characteristics of particle regions using a single scale or a single morphological method alone.
[0004] To address these issues, we devised a cascaded center point fusion method. This method first extracts initial cloud particle centers from processed images of different scales and morphologies, then fuses and projects these initial centers back into the original image space. Finally, a further fusion and optimization of these centers is performed within the original image space. This effectively addresses the shortcomings of single-scale or single-morphology processing. Combined with a deep learning model, this method addresses the inaccurate localization of particle region detection. This multi-stage fusion strategy allows for more comprehensive extraction of cloud particle spatial distribution information, significantly improving the stability and accuracy of cloud particle region detection and providing valuable support for subsequent particle classification and identification, as well as cloud microphysics research. Summary of the Invention
[0005] In view of the above-mentioned deficiencies in the prior art, the present invention proposes a cascade center point fusion 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; 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 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: Filter out abnormal center points of the projections within each maximum connected region in the original cloud particle image data; Step 3.3: Fuse the retained center points of the projections within each maximum connected region in the original cloud particle image data; Step 4: Generate an anchor frame using the fusion center point in the original cloud particle image data as the center point, specifically: Step 4.1: Take the fusion center point in the original cloud particle image data as the anchor box center point, and generate T anchor boxes for each anchor box center point. The anchor box generated by each center point uses T1 aspect ratios and T2 scales; 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 the anchor boxes according to their overlap and category likelihood to obtain the 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, step 3.2: performing abnormal center point filtering on the center point of the projection within each maximum connected area in the original cloud particle image data, specifically: calculating the Euclidean distance between all center points within each maximum connected area, and then calculating the average distance between each center point and other center points, using the average distance as a judgment basis, using the outlier method to screen out center points with abnormal distances from other center points, and then removing these center points.
[0011] Furthermore, the step 3.3: fusing the retained center points of the projections within each largest connected region in the original cloud particle image data, specifically: first calculating the number of pixels within 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 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.
[0012] 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; Furthermore, the step 5: establishing a deep learning model to predict the category probability, center point offset and anchor box 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 box size offset of each anchor box; Furthermore, step 6: filtering anchor frames according to their overlap and category likelihood to obtain detection results, specifically: assigning a category confidence score to each generated anchor frame through the classification layer, sorting all anchor frames from high to low according to their category confidence scores, applying the non-maximum suppression method based on the intersection-union ratio of the anchor frames and the category confidence scores, filtering out redundant anchor frames and removing them to obtain the final detection results.
[0013] Compared with the traditional cloud particle target detection method, the present invention has the following advantages, thereby solving the corresponding technical problems: 1. The present invention performs different levels of erosion and dilation operations on the original cloud particle image data, so that large-sized cloud particles in the image are gradually highlighted through multiple erosion processes, while ensuring that small-sized cloud particles are not missed through multiple dilation processes, thereby effectively improving the adaptability of cloud particle target detection in complex scenes with obvious scale differences.
[0014] 2. The present invention extracts multiple center points, including the geometric center point, the center point of the minimum circumscribed rectangle, the center point of the maximum connected area, the contour centroid, and the center point of the minimum circumscribed circle, from each image data after undergoing different morphological processing. This effectively enhances the reliability of the center points in expressing the location of the particle region. Furthermore, for the multiple extracted center points, the present invention further introduces an outlier center point filtering mechanism based on the distance between center points. This mechanism uses an outlier method to remove outliers with significant positional deviations, significantly reducing the error in center point positioning.
[0015] 3. The present invention proposes a cascade center point fusion strategy to further improve the center point positioning accuracy and reduce the amount of calculation. First, in each image after different levels of morphological processing, a fusion threshold corresponding to the variable is set according to the number of pixels in the connected area. The distance between all center points in each area is calculated. If the distance between the center points is less than the threshold, these center points are fused for the first time. Subsequently, the center points after the first fusion are uniformly projected back to the original cloud particle image data coordinate space for a second cyclic fusion until the center points no longer meet the fusion conditions. This cascade fusion strategy can correct for minor positioning deviations that may be introduced by multi-scale morphological processing and coordinate projection processes by first performing preliminary fusion in the image after morphological processing and then performing center point fusion again in the original image coordinate space, further ensuring the accuracy and stability of the final center point position.
[0016] 4. This paper designs a mechanism for standardizing and filtering the number of anchor frames suitable for cloud particle detection. A preset number of anchor frames is generated based on the fused center point. Anchor frame filtering criteria are designed based on scale and aspect ratio. By controlling overlap between anchor frames, redundant anchor frames are effectively removed, improving the efficiency of integration with deep learning models and the robustness of detection results. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] Figure 1 This is a flow chart of a cascade center point fusion cloud particle detection method. DETAILED DESCRIPTION
[0018] 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; 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 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: Filter out abnormal center points of the projections within each maximum connected region in the original cloud particle image data; Step 3.3: Fuse the retained center points of the projections within each maximum connected region in the original cloud particle image data; Step 4: Generate an anchor frame using the fusion center point in the original cloud particle image data as the center point, specifically: Step 4.1: Take the fusion center point in the original cloud particle image data as the anchor box center point, and generate T anchor boxes for each anchor box center point. The anchor box generated by each center point uses T1 aspect ratios and T2 scales; 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 the anchor boxes according to their overlap and category likelihood to obtain the detection results.
[0019] 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.
[0020] 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.
[0021] 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.
[0022] 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.
[0023] Furthermore, step 3.2: performing abnormal center point filtering on the center point of the projection within each maximum connected area in the original cloud particle image data, specifically: calculating the Euclidean distance between all center points within each maximum connected area, and then calculating the average distance between each center point and other center points, using the average distance as a judgment basis, using the outlier method to screen out center points with abnormal distances from other center points, and then removing these center points.
[0024] Furthermore, the step 3.3: fusing the retained center points of the projections within each largest connected region in the original cloud particle image data, specifically: first calculating the number of pixels within 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 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, 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; Furthermore, the step 5: establishing a deep learning model to predict the category probability, center point offset and anchor box 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 box size offset of each anchor box; Furthermore, step 6: filtering anchor frames according to their overlap and category likelihood to obtain detection results, specifically: assigning a category confidence score to each generated anchor frame through the classification layer, sorting all anchor frames from high to low according to their category confidence scores, applying the non-maximum suppression method based on the intersection-union ratio of the anchor frames and the category confidence scores, filtering out redundant anchor frames and removing them to obtain the final detection results.
[0026] 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 cascade center point fusion 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; 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 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: Filter out abnormal center points of the projections within each maximum connected region in the original cloud particle image data; Step 3.3: Fuse the retained center points of the projections within each maximum connected region in the original cloud particle image data; Step 4: Generate an anchor frame using the fusion center point in the original cloud particle image data as the center point, specifically: Step 4.1: Take the fusion center point in the original cloud particle image data as the anchor box center point, and generate T anchor boxes for each anchor box center point. The anchor box generated by each center point uses T1 aspect ratios and T2 scales; 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 the anchor boxes according to their overlap and category likelihood to obtain the detection results.
2. A cascade center point fusion 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 cascade center point fusion 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 cascade center point fusion 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 screen out center points with abnormal distances from other center points, and then these center points are removed.
5. A cascade center point fusion 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 cascade center point fusion cloud particle detection method as claimed in claim 1, characterized in that The step 3.2: performing abnormal center point filtering on the center point of the projection within each maximum connected area in the original cloud particle image data, specifically, calculating the Euclidean distance between all center points within each maximum connected area, and then calculating the average distance between each center point and other center points, using the average distance as a judgment basis, using the outlier method to filter out center points with abnormal distances from other center points, and then removing these center points.
7. A cascade center point fusion cloud particle detection method as claimed in claim 1, characterized in that Step 3.3: Fusing the retained center points of the projections within each maximum connected region in the original cloud particle image data, specifically: first calculating the number of pixels within the maximum connected region; then, setting an adaptive threshold based on the number of pixels in the maximum 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 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.
8. A cascade center point fusion 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.
9. A cascade center point fusion cloud particle detection method as claimed in 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.
10. A cascade center point fusion cloud particle detection method as claimed in claim 1, characterized in that The step 6: filtering anchor frames according to the overlap and category possibility of the anchor frames to obtain the detection results, specifically: assigning a category confidence score to each generated anchor frame through the classification layer, sorting all anchor frames from high to low according to the category confidence scores, applying the non-maximum suppression method according to the intersection-union ratio of the anchor frames and the category confidence scores, filtering out redundant anchor frames and removing them to obtain the final detection results.