Dynamic label distribution strategy fused with anchor-free strategy
By adopting anchor-free strategy and TaskAlignedAssigner allocation strategy in dynamic image processing, the anchor strategy parameter setting difficulties in dynamic image processing and the IOU matching strategy are solved, and more efficient object detection and segmentation performance are achieved.
Patent Information
- Application Number
- CN202411327118.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-23
- Publication Date
- 2025-01-07
AI Technical Summary
In the prior art In dynamic image processing, anchor strategy requires a large number of parameter settings and is not suitable for different objects. The IOU matching strategy is sensitive to target size and shape, resulting in poor target detection and segmentation performance.
A dynamic tag allocation strategy that integrates anchor-free strategy is adopted, and through equivalent copying of video images, dynamic fuzzing processing and TaskAlignedAssigner allocation strategy, flexible matching and tagging of targets is achieved, replacing the traditional anchor and IOU matching strategies.
A large number of prior parameter settings are avoided, the target detection accuracy is improved, the DFL regression loss problem is solved, and the target detection and segmentation performance is improved.
Smart Images
Figure CN119274208A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of image processing technology, and in particular to a dynamic label allocation strategy integrating an anchor-free strategy. Background Art
[0002] For the processing of dynamic images, the existing methods generally adopt a single anchor strategy or an IOU matching strategy. The setting of the anchor strategy requires setting a large number of parameters, and the discrete anchor scale setting may cause some objects to not be well matched to the anchor, resulting in omissions and increased subsequent DFL losses. The IOU matching method is very sensitive to the target size and the target shape, which in turn affects the deviation. It also cannot completely match the target shape with the bounding box. IoU may not be able to accurately evaluate the performance of target detection or segmentation.
[0003] Therefore, it is necessary to design a dynamic label allocation strategy that integrates the anchor-free strategy. Summary of the invention
[0004] The purpose of the present invention is to provide a dynamic label allocation strategy that integrates an anchor-free strategy to solve the problems raised in the above background technology.
[0005] In order to solve the above technical problems, the present invention provides the following technical solutions:
[0006] A dynamic label allocation strategy integrating an anchor-free strategy includes an anchor-free strategy, a dynamic fuzzy logic part and a TaskAlignedAssigner allocation strategy, which includes the following steps:
[0007] S100, importing the video image into the image detection system built based on YOLO to classify and detect the video image;
[0008] S200, during the detection process, the video image is equivalently copied to form two equal-quality data streams, and then two processings based on center area prediction and key point prediction are performed through an anchor-free strategy;
[0009] S300, importing the two data streams processed by S200 into the dynamic fuzzy logic unit for dynamic fuzzy processing;
[0010] S400, merging the two pieces of data after dynamic fuzzy processing in S300, removing the non-overlapping parts, and emphasizing the overlapping parts;
[0011] S500, dynamically label the emphasized image generated in S400 through the TaskAlignedAssigner allocation strategy to achieve video imaging.
[0012] According to the above technical solution, the specific method of S100 is as follows:
[0013] After the video image is imported into the image detection system, the human body parts in the image are effectively segmented through image segmentation. After segmentation, the image fragments are screened and classified again to form video data streams of different body parts.
[0014] According to the above technical solution, the specific method of S200 is as follows:
[0015] Through multi-node replication correction, the video is streamed and replicated at the same time to form two equal-quality data streams. The two equal-quality data streams are processed by anchor-free strategies respectively, one based on central area prediction and the other based on key point estimation:
[0016] Based on key point estimation: Target detection is performed through a pair of diagonal points. A pair of diagonal points of the same target has high similarity, while a pair of diagonal points of different targets has a large distance.
[0017] Prediction based on the central area: By detecting the distance between the point and the four true sides, all points in the true box are positively sampled, and the level is determined by the maximum value of the distance from the point to the four sides.
[0018] According to the above technical solution, the specific method of S300 is as follows:
[0019] The two data streams processed by S200 enter the dynamic fuzzy logic unit and undergo dynamic fuzzy processing respectively:
[0020] After the data stream processed by the key point estimation strategy is subjected to dynamic blur processing, the first characteristic dynamic image is formed;
[0021] After dynamic blur processing is performed on the data stream processed based on the central area prediction strategy, a second characteristic dynamic image is formed.
[0022] According to the above technical solution, the specific method of S400 is as follows:
[0023] The first characteristic dynamic image generated by S300 is overlapped with the second characteristic dynamic image, the overlapping parts are emphasized, and the non-overlapping parts are removed, thereby forming a fused comprehensive characteristic dynamic image.
[0024] According to the above technical solution, the specific method of S500 is:
[0025] Through the TaskAlignedAssigner allocation strategy, the comprehensive feature dynamic images are flexibly labeled based on content features. Through specific label marking, the dynamic images are cut into several single static feature images, and then the corresponding single convolution training is performed through the YOLOv8 neural network to achieve effective allocation.
[0026] According to the above technical solution, the specific method of image segmentation is to construct a feature pyramid through FPN, effectively and quickly locate and detect the body part information in the picture, divide the boundaries of the body parts based on the feature points in the image, and then copy and extend the image to both sides based on the boundaries, so as to ensure the integrity of the image after cutting.
[0027] According to the above technical solution, the method for duplicating the same-quality data stream of the video film data streaming is specifically as follows:
[0028] The data stream is effectively node-cut, that is, based on the complete body parts as nodes, the data stream segments are formed, and the feature data in the data stream segments are then combined in pairs to form feature data groups. The two groups of feature data are divided based on the binary method, and then copied separately, and then the copied feature data are fitted and corrected with the original feature data, thereby forming two equal data streams.
[0029] According to the above technical solution, the method for superimposing the first characteristic dynamic image and the second characteristic dynamic image is as follows:
[0030] The first feature dynamic image and the second feature dynamic image are subjected to gradient depth projection, and overlapping projection points are obtained based on different depths. Feature points are then reversely searched based on the overlapping projection points to obtain areas with different overlaps, and a fused feature dynamic image is constructed through the overlap areas.
[0031] According to the above technical solution, the cutting of the dynamic image is effectively segmented based on the label marking, that is, the current label is the starting point and the subsequent labels are the ending points.
[0032] Compared with the prior art, the beneficial effects achieved by the present invention are:
[0033] By replacing the anchor strategy with the anchor-free strategy, a large number of prior parameter settings can be avoided, which can ensure that any target object matches the anchor-free strategy and effectively solve the DFL regression loss problem. At the same time, by replacing the IOU matching strategy with the TaskAlignedAssigner dynamic label allocation strategy, the target detection accuracy can be improved. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] Figure 1 It is a flowchart of the present invention. DETAILED DESCRIPTION
[0035] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.
[0036] like Figure 1 As shown, the present invention provides a technical solution: a dynamic label allocation strategy integrating an anchor-free strategy, including an anchor-free strategy, a dynamic fuzzy logic unit and a TaskAlignedAssigner allocation strategy, which contains the following steps:
[0037] S100, importing the video image into the image detection system built based on YOLO to classify and detect the video image;
[0038] S200, during the detection process, the video image is equivalently copied to form two equal-quality data streams, and then two processings based on center area prediction and key point prediction are performed through an anchor-free strategy;
[0039] S300, importing the two data streams processed by S200 into the dynamic fuzzy logic unit for dynamic fuzzy processing;
[0040] S400, merging the two pieces of data after dynamic fuzzy processing in S300, removing the non-overlapping parts, and emphasizing the overlapping parts;
[0041] S500, dynamically label the emphasized image generated in S400 through the TaskAlignedAssigner allocation strategy to achieve video imaging.
[0042] Specifically, the specific methods of S100 are as follows:
[0043] After the video image is imported into the image detection system, the human body parts in the image are effectively segmented through image segmentation. After segmentation, the image fragments are screened and classified again to form video data streams of different body parts.
[0044] Specifically, the specific methods of S200 are as follows:
[0045] Through multi-node replication correction, the video is streamed and replicated at the same time to form two equal-quality data streams. The two equal-quality data streams are processed by anchor-free strategies respectively, one based on central area prediction and the other based on key point estimation:
[0046] Based on key point estimation: Target detection is performed through a pair of diagonal points. A pair of diagonal points of the same target has high similarity, while a pair of diagonal points of different targets has a large distance.
[0047] Prediction based on the central area: By detecting the distance between the point and the four true sides, all points in the true box are positively sampled, and the level is determined by the maximum value of the distance from the point to the four sides.
[0048] Specifically, the specific methods of S300 are as follows:
[0049] The two data streams processed by S200 enter the dynamic fuzzy logic unit and undergo dynamic fuzzy processing respectively:
[0050] After the data stream processed by the key point estimation strategy is subjected to dynamic blur processing, the first characteristic dynamic image is formed;
[0051] After dynamic blur processing is performed on the data stream processed based on the central area prediction strategy, a second characteristic dynamic image is formed.
[0052] Specifically, the specific methods of S400 are as follows:
[0053] The first characteristic dynamic image generated by S300 is overlapped with the second characteristic dynamic image, the overlapping parts are emphasized, and the non-overlapping parts are removed, thereby forming a fused comprehensive characteristic dynamic image.
[0054] Specifically, the specific methods of S500 are:
[0055] Through the TaskAlignedAssigner allocation strategy, the comprehensive feature dynamic images are flexibly labeled based on content features. Through specific label marking, the dynamic images are cut into several single static feature images, and then the corresponding single convolution training is performed through the YOLOv8 neural network to achieve effective allocation.
[0056] Specifically, the specific method of image segmentation is to construct a feature pyramid through FPN, effectively and quickly locate and detect the body part information in the picture, divide the boundaries of the body parts based on the feature points in the image, and then copy and extend the image to both sides based on the boundaries, so as to ensure the integrity of the image after cutting.
[0057] Specifically, the video data streaming isochronous data stream replication method is as follows:
[0058] The data stream is effectively node-cut, that is, based on the complete body parts as nodes, the data stream segments are formed, and the feature data in the data stream segments are then combined in pairs to form feature data groups. The two groups of feature data are divided based on the binary method, and then copied separately, and then the copied feature data are fitted and corrected with the original feature data, thereby forming two equal data streams.
[0059] Specifically, the method for superimposing the first characteristic dynamic image and the second characteristic dynamic image is as follows:
[0060] The first feature dynamic image and the second feature dynamic image are subjected to gradient depth projection, and overlapping projection points are obtained based on different depths. Feature points are then reversely searched based on the overlapping projection points to obtain areas with different overlaps, and a fused feature dynamic image is constructed through the overlap areas.
[0061] Specifically, the cutting of the dynamic image is based on label marking to perform effective segmentation, that is, the current label is the starting point and the subsequent labels are the ending points.
[0062] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.
[0063] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.
Claims
1. A dynamic label allocation strategy integrating an anchor-free strategy, including an anchor-free strategy, a dynamic fuzzy logic unit and a TaskAlignedAssigner allocation strategy, characterized in that: It contains the following steps: S100, importing the video image into the image detection system built based on YOLO to classify and detect the video image; S200, during the detection process, the video image is equivalently copied to form two equal-quality data streams, and then two processings based on center area prediction and key point prediction are performed through an anchor-free strategy; S300, importing the two data streams processed by S200 into the dynamic fuzzy logic unit for dynamic fuzzy processing; S400, merging the two pieces of data after dynamic fuzzy processing in S300, removing the non-overlapping parts, and emphasizing the overlapping parts; S500, dynamically label the emphasized image generated in S400 through the TaskAlignedAssigner allocation strategy to achieve video imaging.
2. According to claim 1, a dynamic label allocation strategy integrating an anchor-free strategy is characterized in that: The specific method of S100 is as follows: After the video image is imported into the image detection system, the human body parts in the image are effectively segmented through image segmentation. After segmentation, the image fragments are screened and classified again to form video data streams of different body parts.
3. According to claim 1, a dynamic label allocation strategy integrating an anchor-free strategy is characterized in that: The specific method of S200 is as follows: Through multi-node replication correction, the video is streamed and replicated at the same time to form two equal-quality data streams. The two equal-quality data streams are processed by anchor-free strategies respectively, one based on central area prediction and the other based on key point estimation: Based on key point estimation: Target detection is performed through a pair of diagonal points. A pair of diagonal points of the same target has high similarity, while a pair of diagonal points of different targets has a large distance. Prediction based on the central area: By detecting the distance between the point and the four true sides, all points in the true box are positively sampled, and the level is determined by the maximum value of the distance from the point to the four sides.
4. According to claim 1, a dynamic label allocation strategy integrating an anchor-free strategy is characterized in that: The specific method of S300 is as follows: The two data streams processed by S200 enter the dynamic fuzzy logic unit and undergo dynamic fuzzy processing respectively: After the data stream processed by the key point estimation strategy is subjected to dynamic blur processing, the first characteristic dynamic image is formed; After dynamic blur processing is performed on the data stream processed based on the central area prediction strategy, a second characteristic dynamic image is formed.
5. According to claim 1, a dynamic label allocation strategy integrating an anchor-free strategy is characterized in that: The specific method of S400 is as follows: The first characteristic dynamic image generated by S300 is overlapped with the second characteristic dynamic image, the overlapping parts are emphasized, and the non-overlapping parts are removed, thereby forming a fused comprehensive characteristic dynamic image.
6. According to claim 1, a dynamic label allocation strategy integrating an anchor-free strategy is characterized in that: The specific method of S500 is: Through the TaskAlignedAssigner allocation strategy, the comprehensive feature dynamic images are flexibly labeled based on content features. Through specific label marking, the dynamic images are cut into several single static feature images, and then the corresponding single convolution training is performed through the YOLOv8 neural network to achieve effective allocation.
7. The dynamic label allocation strategy integrating the anchor-free strategy according to claim 2, characterized in that: The specific method of image segmentation is to construct a feature pyramid through FPN, effectively and quickly locate and detect the body part information in the image, divide the boundaries of the body parts based on the feature points in the image, and then copy and extend the image to both sides based on the boundaries, so as to ensure the integrity of the image after cutting.
8. According to claim 3, a dynamic label allocation strategy integrating an anchor-free strategy is characterized in that: The method for duplicating the same-quality data stream of the video film and television data streaming is specifically as follows: The data stream is effectively node-cut, that is, based on the complete body parts as nodes, the data stream segments are formed, and the feature data in the data stream segments are then combined in pairs to form feature data groups. The two groups of feature data are divided based on the binary method, and then copied separately, and then the copied feature data are fitted and corrected with the original feature data, thereby forming two equal data streams.
9. The dynamic label allocation strategy integrating the anchor-free strategy according to claim 5, characterized in that: The method for superimposing the first characteristic dynamic image and the second characteristic dynamic image is as follows: The first feature dynamic image and the second feature dynamic image are subjected to gradient depth projection, and overlapping projection points are obtained based on different depths. Feature points are then reversely searched based on the overlapping projection points to obtain areas with different overlaps, and a fused feature dynamic image is constructed through the overlap areas.
10. The dynamic label allocation strategy integrating the anchor-free strategy according to claim 6, characterized in that: The cutting of the dynamic image is based on label marking to perform effective segmentation, that is, the current label is the starting point and the subsequent labels are the end points.