Multi-target detection method based on dynamic adjustment and frequency-channel double cooperation
By adopting a multi-target detection method with dynamic adjustment and frequency-channel dual coordination, combined with YOLOv8 and DFCC-Mask R-CNN, the problem of balancing accuracy and efficiency in complex scenarios is solved, achieving highly adaptive multi-target detection and improving the reliability and real-time performance of detection results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-03
- Publication Date
- 2026-04-10
AI Technical Summary
Existing target detection algorithms struggle to balance accuracy and efficiency in complex scenarios, lack dynamic adaptability, and fail to meet the real-time requirements of multi-target detection.
A multi-target detection method based on dynamic adjustment and frequency-channel dual collaboration is designed. Through a dynamic hierarchical inference mechanism, YOLOv8, DFCC-Mask R-CNN and Mask R-CNN are combined to dynamically select the detection mode according to the scene complexity. The feature map is optimized by the frequency-channel dual collaborative lightweight operator DFCC to achieve a balance between accuracy and efficiency.
It achieves strong adaptability in different scenarios, and can reduce computing costs while ensuring real-time performance and high accuracy, thereby improving the reliability and adaptability of detection results.
Smart Images
Figure CN121837592A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the field of target detection, and relates to a multi-target detection technology, in particular to a multi-target detection method based on dynamic adjustment and frequency-channel double coordination. BACKGROUND
[0002] With the continuous development of computer vision, deep learning in the field of target detection in visual processing is also continuously expanding. Accurate and efficient target detection technology is the core of realizing intelligent monitoring, and its performance directly affects the intelligent level of the monitoring system. However, the current target detection algorithm still has problems such as large differences in scene complexity, difficulty in balancing precision and efficiency, and insufficient dynamic adaptability in actual scenes.
[0003] At present, the mainstream target detection method mainly includes two types: one is a single-stage detection algorithm represented by YOLO, which has the characteristics of fast speed and good real-time performance, but the detection accuracy in complex scenes is insufficient in details; the other is a two-stage detection algorithm represented by Mask R-CNN, which can realize target detection and has high accuracy, but the algorithm has large computational complexity and slow inference speed, which is difficult to meet the demand of large-scale real-time detection. SUMMARY
[0004] The purpose of the application is to solve the problems of large differences in scene complexity, difficulty in balancing precision and efficiency, and insufficient dynamic adaptability of the current target detection algorithm in actual scenes, and to provide a multi-target detection method based on dynamic adjustment and frequency-channel double coordination, which fully considers the characteristics of large differences in environmental scene complexity, designs a dynamic hierarchical inference mechanism to balance precision and efficiency, and uses a multi-target detection model based on dynamic adjustment and frequency-channel double coordination to dynamically adjust the threshold value of historical inference error to adapt to different detection environments.
[0005] Technical scheme: In order to achieve the above purpose, the application provides a multi-target detection method based on dynamic adjustment and frequency-channel double coordination, which includes the following steps:
[0006] S1: input an image and convert the image format, and use YOLOv8 for preliminary detection;
[0007] S2: based on the preliminary detection result, calculate the target density and background entropy, and obtain the final scene complexity score by proportionally weighting the sum of the two;
[0008] S3: compare the scene complexity score with the scene complexity threshold value to obtain the scene type; the scene type includes simple, medium and complex scenes;
[0009] S4: for a simple scene, detect the scene image by YOLOv8;
[0010] For medium scenes, the constructed DFCC-Mask R-CNN model is used for lightweight detection of scene images;
[0011] For complex scenes, Mask R-CNN is called for full detection;
[0012] S5: Compare the results obtained after step S4 detection with the provided labeled data to obtain the error rate, and dynamically adjust the scene complexity threshold according to the average error;
[0013] S6: The detection results of step S4 are presented as visual results.
[0014] Further, the step S1 comprises:
[0015] A1: Convert the BRG model of the image to the RBG model, and convert the B, R, G parameters to R, B, G parameters by index rearrangement;
[0016] A2: Call the YOLOv8 model to infer the image and obtain the preliminary detection result.
[0017] Further, the step S2 comprises:
[0018] B1: Calculate the target density, and the specific calculation formula is as follows:
[0019]
[0020] Wherein, is the number of targets with confidence higher than the threshold, is the image height, is the image width;
[0021] B2: Calculate the background entropy, and the specific calculation formula is as follows
[0022]
[0023] Wherein, is the probability that the pixel value in the gray image is is the minimum value, is the normalized background entropy, is the background entropy;
[0024] B3: Calculate the scene complexity, and the specific calculation formula is as follows:
[0025]
[0026] Wherein, is the scene complexity.
[0027] Further, the determination method of the scene type in step S3 comprises:
[0028] scene complexity defined as a simple scene, defined as a medium scene, defined as a complex scene.
[0029] Further, the DFCC-Mask R-CNN model in step S4 comprises a backbone network reformed by a lightweight network MobileNetV3-Large and a DFCC feature optimization module, and the DFCC feature optimization module is responsible for redundant filtering and effective information strengthening of the feature map output by the backbone network.
[0030] Further, the DFCC feature optimization module comprises a frequency coordination sub-module and a channel coordination sub-module.
[0031] Further, the operation process of the DFCC feature optimization module in the DFCC-Mask R-CNN model of step S4 comprises:
[0032] C1: frequency decoupling of the input feature tensor along the spatial frequency axis, and the specific calculation formula is as follows:
[0033]
[0034]
[0035] wherein X is the current layer feature map, L is a low-frequency smoothing component, H is a high-frequency residual component, is an average pooling, is the side length of the pooling window, is the step length of each sliding window, is the number of zero pixels supplemented at the input edge, Upsample is the up-sampling of L after pooling back to the original size, and subscript x1 is a shorthand for "magnification = 1";
[0036] C2: applying a learnable gate to the low-frequency branch to generate a channel-level 0-1 mask, and the specific calculation formula is as follows:
[0037]
[0038]
[0039] wherein, is a point-by-point grouping convolution kernel, and s is a channel gating coefficient, is the sparse version of the original component L filtered by the gating mask, is the global average pooling, is the Sigmoid function, is the Hadamard product;
[0040] C3: Energy screening in the frequency domain is performed on the high-frequency branch, and the specific calculation formula is as follows:
[0041]
[0042]
[0043]
[0044]
[0045]
[0046] where F is the complex domain of 2D discrete Fourier transform, E is the frequency energy spectrum, is the learnable threshold coefficient, is the actual energy threshold, M is the binary energy mask, is the Fourier energy gate output, which is the sparse version filtered by the gating mask, FFT is the two-dimensional discrete Fourier transform, IFFT is the two-dimensional inverse discrete Fourier transform, and mean is the global average, is the Hadamard product;
[0047] C4: Group convolution is performed on the gated coefficient features, and the contributions of the two paths are adjusted in real time through learnable weights, and the specific calculation formula is as follows:
[0048]
[0049]
[0050] where a and b are learnable scalars, and a and b are dynamic fusion weights, is the group convolution, is the normalization function, Y is the final output of the DFCC operator, and the shape is the same as X.
[0051] Further, the step S4 directly filters low-confidence YOLOv8 results with simple scenes without calling MaskR-CNN; calls DFCC-Mask R-CNN model without processing masks for medium scenes; and fuses the bounding box and the mask to retain complete semantic information for complex scenes.
[0052] Further, the step S5 includes:
[0053] D1: Determine whether each prediction result matches the labeled data by using the intersection over union (IOU); if no labeled data is provided, skip this step; the specific formula is as follows:
[0054] (15)
[0055] D2: Count the number of successfully matched targets, and calculate the error rate, the specific formula is as follows:
[0056] (16)
[0057] Wherein, |P| is the total number of predicted targets, and |G| is the total number of labeled targets;
[0058] D3: Store the scene complexity and the error rate of the current sample in the buffer, and delete the earliest record when the size of the buffer exceeds the feedback window;
[0059] D4: When the data amount of the buffer reaches the set number of records, divide the scene according to the current threshold value, and calculate the average error of simple scenes and medium scenes:
[0060] (17)
[0061] (18)
[0062] Wherein, is the error rate of a single sample, is the average error of simple scenes, is the average error of medium scenes;
[0063] D5: Dynamically adjust the scene complexity threshold value, when the average error of simple scenes exceeds 20% (0.2), reduce the threshold value, the specific formula is as follows:
[0064] (19)
[0065] Wherein, is the dynamically adjusted simple scene threshold value, is the original simple scene threshold value;
[0066] When the medium scene average error exceeds 15% ( >0.15), the threshold is reduced, and the specific formula is as follows:
[0067] (20)
[0068] Wherein, is the medium scene threshold after dynamic adjustment, is the original medium scene threshold.
[0069] Further, in the step S6: the boundary box is drawn according to the color of the target source, the category name, the confidence and the weight information of the fusion result are labeled, the boundary box color of the YOLO inference mode is blue, the boundary box color of the frequency-channel double coordination lightweight mode is orange, and the boundary box color of the full fusion mode is red; in the full fusion mode, the preset bright color group is used in a loop to superimpose the mask and adjust the weight proportion of the image area and the mask color, and the visual result is saved.
[0070] The present application aims at the problem that the target detection technology is difficult to balance the multi-target detection precision and speed, proposes a multi-target detection model based on dynamic adjustment and frequency-channel double coordination, designs a dynamic adjustment hierarchical inference mechanism, which can meet the real-time requirement of the environment, maintain high detection precision in complex environment, and has certain adaptability and deployment flexibility.
[0071] Advantages: compared with the prior art, the present application has the following advantages:
[0072] (1) The present application can adapt to scene complexity, balance efficiency and precision, and solve the problem that the traditional single model is difficult to balance precision and speed.
[0073] (2) The present application introduces a frequency-channel double coordination lightweight operator (DFCC), proposes a new DFCC-Mask R-CNN, and improves the reliability of the detection result.
[0074] (3) The present application has feedback learning ability, and the model can optimize the threshold autonomously to adapt to multi-target detection tasks in different environments.
[0075] (4) The lightweight fusion strategy is implemented in the medium scene to reduce the calculation cost while ensuring certain precision. DETAILED DESCRIPTION
[0076] Figure 1 is a flowchart of the method of the present application;
[0077] Figure 2is a comparison diagram of the original Mask R-CNN model structure and the DFCC-Mask R-CNN model structure;
[0078] Figure 3 is a typical test image in a complex scene provided by an embodiment of the application;
[0079] Figure 4 is a target detection result in a complex scene provided by an embodiment of the application; Figure 3
[0080] Figure 5 is a typical test image in a medium scene provided by an embodiment of the application;
[0081] Figure 6 is a target detection result in a medium scene provided by an embodiment of the application; Figure 5
[0082] Figure 7 is a typical test image in a simple scene provided by an embodiment of the application;
[0083] Figure 8 is a target detection result in a simple scene provided by an embodiment of the application; Figure 7
[0084] Figure 9 is a performance result of testing only using the DFCC-Mask R-CNN model;
[0085] Figure 10 is a performance result of testing only using the Mask R-CNN model;
[0086] Figure 11 is a performance result of testing only using the YOLOv8 model. DETAILED DESCRIPTION
[0087] The application will be further clarified by the following figures and embodiments, which should be understood as merely illustrating the application and not limiting the scope of the application. After reading the application, those skilled in the art can make various modifications to the application, which fall within the scope of the appended claims.
[0088] Embodiment 1:
[0089] As shown in Figure 1 , the embodiment provides a multi-target detection method based on dynamic adjustment and frequency-channel double coordination, including the following steps:
[0090] S1: input an image and convert the image format, and perform preliminary detection with YOLOv8;
[0091] Step S1 includes:
[0092] A1: Convert the BRG model of the image to the RBG model, and convert the B, R, G parameters to R, B, G parameters by index rearrangement;
[0093] A2: Call the YOLOv8 model to infer the image and obtain the preliminary detection result.
[0094] S2: Based on the preliminary detection result, calculate the target density and background entropy, and obtain the final scene complexity score by weighted sum of the two;
[0095] Step S2 includes:
[0096] B1: Calculate the target density, and the specific calculation formula is as follows:
[0097]
[0098] wherein, is the number of targets with confidence higher than the threshold, is the image height, is the image width;
[0099] B2: Calculate the background entropy, and the specific calculation formula is as follows
[0100]
[0101] wherein, is the probability of the pixel value being in the grayscale image, is the minimum value, is the normalized background entropy, is the background entropy;
[0102] B3: Calculate the scene complexity, and the specific calculation formula is as follows:
[0103]
[0104] wherein, is the scene complexity.
[0105] S3: Compare the scene complexity score with the scene complexity threshold to obtain the scene type; the scene type includes simple, medium and complex scenes;
[0106] Perform hierarchical inference process and dynamically select mode according to scene complexity;
[0107] Scene complexity is defined as a simple scene, is defined as a medium scene, is defined as a complex scene.
[0108] S4: For a simple scene, the scene image is detected by YOLOv8;
[0109] For a medium scene, a Dual-Frequency-Channel Collaborative (DFCC) lightweight operator is introduced to construct a DFCC-Mask R-CNN model, and the scene image is detected by the constructed DFCC-Mask R-CNN model;
[0110] For a complex scene, Mask R-CNN is called for full detection;
[0111] As shown in Figure 2 Compared with the original Mask R-CNN model structure, the DFCC-Mask R-CNN model includes a lightweight network MobileNetV3-Large lightweight transformation backbone network, and the parameter quantity is reduced to 1 / 5 of the original ResNet-50; wherein the DFCC feature optimization module is responsible for filtering the feature map output by the backbone network and strengthening the effective information, which is specifically composed of a frequency collaborative submodule and a channel collaborative submodule, and the FPN channel number of each layer is reduced from 256 dimensions to 128 dimensions lightweight FPN.
[0112] The running process of the DFCC feature optimization module in the DFCC-Mask R-CNN model includes:
[0113] C1: Frequency decoupling of the input feature tensor along the spatial frequency axis, the specific calculation formula is as follows:
[0114]
[0115]
[0116] Wherein, X is the current layer feature map, L is the low-frequency smooth component, H is the high-frequency residual component, is the average pooling, is the side length of the pooling window, is the step length of each sliding window, is the number of zero pixels supplemented at the input edge, Upsample is to upsample L after pooling back to the original size, and subscript x1 is a shorthand for "magnification = 1";
[0117] C2: Learning gate is applied to the low-frequency branch to generate a channel-level 0-1 mask, and the specific calculation formula is as follows:
[0118]
[0119]
[0120] wherein, is the point-wise grouped convolution kernel, s is the channel gating coefficient, is the channel sparse gate output, which is the sparse version of the original component L filtered by the gating mask, is the global average pooling, is the Sigmoid function, is the Hadamard product;
[0121] C3: Energy screening is performed on the high-frequency branch in the frequency domain, and the specific calculation formula is as follows:
[0122]
[0123]
[0124]
[0125]
[0126]
[0127] wherein, F is a complex domain of 2D discrete Fourier transform, E is a frequency energy spectrum, is a learnable threshold coefficient, is an actual energy threshold, M is a binary energy mask, is the Fourier energy gate output, which is the sparse version filtered by the gating mask, FFT is a two-dimensional discrete Fourier transform, IFFT is a two-dimensional inverse discrete Fourier transform, mean is a global average, is the Hadamard product;
[0128] C4: Grouped convolution is performed on the gated coefficient feature, and the contributions of the two paths are adjusted in real time through learnable weights, and the specific calculation formula is as follows:
[0129]
[0130]
[0131] wherein, a, b are learnable scalars, and α, β are dynamic fusion weights, is the grouped convolution, Y is the final output of the DFCC operator, which has the same shape as X.
[0132] In the above formula, formulas (4)-(5) belong to frequency decoupling; formulas (6)-(7) belong to the channel coordination submodule, where the channel gating is applied to the low-frequency component L to generate sparse features L̃; formulas (8)-(14) are the operation of the frequency coordination submodule.
[0133] Simple scenes directly filter low-confidence YOLOv8 results without calling Mask R-CNN; medium scenes call the DFCC-Mask R-CNN model without processing the mask; complex scenes fuse the bounding box and the mask to retain complete semantic information.
[0134] S5: Compare the results obtained after step S4 detection with the provided labeled data to obtain the error rate, and dynamically adjust the scene complexity threshold according to the average error;
[0135] Step S5 includes:
[0136] D1: Determine whether each prediction result matches the labeled data by using the intersection over union (IOU); if no labeled data is provided, skip this step; the specific formula is as follows:
[0137] (15)
[0138] D2: Count the number of successfully matched targets (matched) and calculate the error rate, the specific formula is as follows:
[0139] (16)
[0140] Where |P| is the total number of predicted targets, and |G| is the total number of labeled targets.
[0141] D3: Store the scene complexity and the error rate of the current sample in the buffer, and delete the earliest record when the buffer size exceeds the feedback window (set to 50 in this embodiment);
[0142] D4: When the buffer data reaches 50, divide the scene according to the current threshold and calculate the average error of simple scenes and medium scenes:
[0143] (17)
[0144] (18)
[0145] Where, is the error rate of a single sample, is the average error of simple scenes, is the medium scene average error;
[0146] D5: dynamically adjust the scene complexity threshold, when the simple scene average error exceeds 20% ( > 0.2), reduce the threshold, the specific formula is as follows:
[0147] (19)
[0148] wherein, is the dynamically adjusted simple scene threshold, is the original simple scene threshold;
[0149] when the medium scene average error exceeds 15% ( > 0.15), reduce the threshold, the specific formula is as follows:
[0150] (20)
[0151] wherein, is the dynamically adjusted medium scene threshold, is the original medium scene threshold.
[0152] S6: the detection result of step S4 is presented as a visualization result.
[0153] In this embodiment, the color of the bounding box is distinguished according to the target source, the class name, the confidence and the weight information of the fusion result are marked, the color of the bounding box of the YOLO inference mode is blue, the color of the bounding box of the frequency-channel dual coordination lightweight mode is orange, and the color of the bounding box of the full-amount fusion mode is red; in the full-amount fusion mode, the preset bright color group is used to superimpose the mask and adjust the weight proportion of the image area and the mask color, and the visualization result is saved.
[0154] Embodiment 2:
[0155] In order to verify the effectiveness and effect of the method of the application, the following test experiments and analysis are carried out in this embodiment:
[0156] Figure 3 is a typical test image under a complex scene; Figure 4 is a target detection result image based on Figure 3 exhibited.
[0157] Figure 5 is a typical test image under a medium scene; Figure 6 is a target detection result image based on Figure 5 exhibited.
[0158] Figure 7 is a typical test image under a simple scene;Figure 8 is based on Figure 7 The target detection result image is displayed.
[0159] From Figure 4 It can be seen that the target detection result accuracy in a complex scene can meet the requirements; from Figure 6 It can be seen that the target detection result accuracy in a medium scene can meet the requirements, and the output result of the lightweight detection can clearly show the detection result relative to the complex scene; from Figure 8 It can be seen that there are fewer targets in a simple scene, and the output result of YOLOv8 can meet the required accuracy.
[0160] Because the number of simple scene images in the coco dataset for testing is limited, there are only 440 images in the simple scene, so 440 images are used for testing in the simple scene, and 1000 images are selected for testing in the complex scene and the medium scene. Figure 9 is the test result of the multi-target detection method based on dynamic adjustment and frequency-channel double coordination provided in the embodiment, Figure 10 is the test result of only using the Mask R-CNN model, Figure 11 is the test result of only using the YOLOv8 model.
[0161] Table 1
[0162]
[0163] Table 1 is a performance comparison of various different models. Overall, DFCC-Mask R-CNN has the best overall performance in balancing accuracy and time consumption in three scenes, YOLOV8 has less time consumption but slightly lower accuracy, and Mask R-CNN has slightly better accuracy but slightly higher time consumption. It can be known that, compared with the Mask R-CNN model and the YOLOv8 model, the present application has better balance in accuracy and speed, especially in a variable scene environment, and has better comprehensive performance, and can adapt to multi-target detection in different scene environments.
Claims
1. A multi-target detection method based on dynamic adjustment and frequency-channel dual coordination, characterized in that, Includes the following steps: S1: Input the image and convert its format, then perform preliminary detection using YOLOv8; S2: Based on the preliminary detection results, calculate the target density and background entropy, and sum the two proportionally to obtain the final scene complexity score; S3: Compare the scene complexity score with the scene complexity threshold to determine the scene type; scene types include simple, medium and complex scenes. S4: For simple scenes, YOLOv8 is used to detect scene images; For medium-scale scenarios, a lightweight detection method is used to perform scene image detection using a pre-constructed DFCC-Mask R-CNN model. For complex scenarios, Mask R-CNN is used for full detection; S5: Compare the results obtained after detection in step S4 with the provided labeled data to obtain the error rate, and dynamically adjust the scene complexity threshold based on the average error. S6: Present the detection results from step S4 as a visualization result.
2. The multi-target detection method based on dynamic adjustment and frequency-channel dual coordination according to claim 1, characterized in that, Step S1 includes: A1: Convert the BRG model of the image to the RBG model. The conversion of B, R, G parameters to R, B, G parameters is achieved by index rearrangement. A2: Call the YOLOv8 model to perform inference on the image and obtain preliminary detection results.
3. The multi-target detection method based on dynamic adjustment and frequency-channel dual coordination according to claim 2, characterized in that, Step S2 includes: B1: Calculate the target density. The specific calculation formula is as follows: ； in, The number of targets with a confidence level higher than the threshold. It is the image height. It is the image width; B2: Calculate the background entropy, the specific calculation formula is as follows: ； in, The pixel value in a grayscale image The probability, It is a local minimum. It is the normalized background entropy. It is background entropy; B3: Calculate the scenario complexity, the specific calculation formula is as follows: ； in, It refers to the complexity of the scene.
4. The multi-target detection method based on dynamic adjustment and frequency-channel dual coordination according to claim 3, characterized in that, The method for determining the scene type in step S3 includes: Scene complexity Defined as a simple scenario, Defined as a medium-level scenario, Defined as a complex scenario.
5. A multi-target detection method based on dynamic adjustment and frequency-channel dual coordination according to claim 3, characterized in that, In step S4, the DFCC-Mask R-CNN model includes a backbone network modified from the lightweight MobileNetV3-Large network and a DFCC feature optimization module. The DFCC feature optimization module is responsible for redundancy filtering and effective information enhancement of the feature maps output by the backbone network.
6. The multi-target detection method based on dynamic adjustment and frequency-channel dual coordination according to claim 5, characterized in that, The DFCC feature optimization module includes a frequency coordination submodule and a channel coordination submodule.
7. A multi-target detection method based on dynamic adjustment and frequency-channel dual coordination according to claim 6, characterized in that, The operation process of the DFCC feature optimization module in the DFCC-Mask R-CNN model in step S4. include: C1: Decouple the input feature tensor along the spatial frequency axis. The specific calculation formula is as follows: ； ； Where X is the feature map of the current layer, L is the low-frequency smoothing component, and H is the high-frequency residual component. For average pooling, Let be the side length of the pooling window. Let this be the step size for each window slide. To add zero-pixel layers at the input edges, Upsample upsamples the pooled L back to its original size, and the subscript ×1 is shorthand for "magnification = 1"; C2: Apply learnable gating to the low-frequency branch to generate a channel-level 0-1 mask. The specific calculation formula is as follows: ； ； in, The kernel is a point-by-point grouped convolution kernel, and s is the channel gating coefficient. This is the sparse gate output of the channel, which is the sparse version of the original component L after being filtered by a gated mask. For global average pooling, For the Sigmoid function, For Hadamard product; C3: Perform energy filtering on the high-frequency branch in the frequency domain. The specific calculation formula is as follows: ； ； ； ； ； Where F is the complex domain of the 2D discrete Fourier transform, and E is the frequency domain energy spectrum. For learnable threshold coefficients, M is the actual energy threshold, and M is the binary energy mask. The Fourier energy gate output is a sparse version filtered by a gated mask. FFT is the two-dimensional discrete Fourier transform, IFFT is the two-dimensional inverse discrete Fourier transform, and mean is the global average. For Hadamard product; C4: Performs grouped convolution on the gated coefficient features and adjusts the contributions of the two paths in real time using learnable weights. The specific calculation formula is as follows: ； ； Where a and b are learnable scalars, and α and β are dynamic fusion weights. For grouped convolution, Y is the normalization function, and Y is the final output of the DFCC operator, with the same shape as X.
8. A multi-target detection method based on dynamic adjustment and frequency-channel dual coordination according to claim 7, characterized in that, In step S4, for simple scenarios, low-confidence YOLOv8 results are directly filtered out without calling Mask R-CNN; for medium-level scenarios, the DFCC-Mask R-CNN model is called without processing the mask; for complex scenarios, the bounding boxes and masks are fused to retain complete semantic information.
9. A multi-target detection method based on dynamic adjustment and frequency-channel dual coordination according to claim 8, characterized in that, Step S5 includes: D1: Compare each prediction result with the labeled data and determine if they match using the Intersection over Union (IOU). Skip this step if no labeled data is provided. The specific formula is as follows: (15)； D2: Count the number of all successfully matched targets and calculate the error rate. The specific formula is as follows: (16)； Where |P| represents the total number of predicted targets, and |G| represents the total number of labeled targets; D3: The scene complexity of the current sample. and error rate Store in the buffer; when the buffer size exceeds the feedback window, delete the oldest record. D4: When the buffer data volume reaches the set threshold, divide the scene according to the current threshold and calculate the average error of the simple and medium scenes: (17)； (18)； in, It is the single-sample error rate. It is the average error in a simple scenario. This is the average error for a medium-scale scenario; D5: Dynamically adjust the scene complexity threshold. When the average error of a simple scene exceeds 20%, lower the threshold. The specific formula is as follows: (19)； in, For the dynamically adjusted simple scene threshold, The threshold for the original, simple scene; When the average error in medium-level scenarios exceeds 15%, the threshold is lowered, as shown in the following formula: (20)； in, The threshold for medium-scale scenarios is dynamically adjusted. This is the original medium scene threshold.