High-confidence intelligent detection method and device for small targets with dense clusters, and medium

Abstract

The invention relates to a high-confidence intelligent detection method and device for a cluster-dense small target, and a medium. The method comprises the steps: constructing an HLS-YOLO-P network model which comprises an HWD down-sampling module based on wavelet transformation, a global feature extraction module C2f-LSKA and a task alignment detection head STADH; and performing multi-layer feature fusion and context information enhancement processing on the image data containing the target scene to obtain a detection result containing a target category, a position and confidence. And carrying out post-processing on the detection result, wherein the post-processing comprises intersection-to-union ratio and confidence threshold adjustment. And finally, generating a corresponding inspection report. Compared with the prior art, the method has the advantages of high detection precision, light weight, wide application range and the like.

Description

Technical Field

[0001] The present invention relates to the fields of computer vision and deep learning, and in particular to a high-confidence intelligent detection method, device and medium for densely clustered small targets. Background Art

[0002] In the field of electric power visual inspection and 3D reconstruction, limited by observation distance, viewing angle, camera pixels and other conditions, such as in the application of photovoltaic and wind turbine fine hardware parts, key components and devices of equipment, power cabinets and data center status lights, the phenomenon of absolutely small and relatively small targets in the image samples obtained by existing machine vision is extremely common. In land-based and air-based inspection and electric power remote vision detection images, due to the insufficient number of pixels, it is impossible to provide texture, boundary, morphology and other identification features in the form of redundant pixels like large targets, resulting in neural network convolution and pooling operations to misunderstand and fail to transmit target features, wasting the already scarce limited information, thus causing distortion, failure, error and other unfavorable situations in electric power visual inspection and 3D reconstruction tasks. Especially in the task of detecting small targets with high density and high similarity, the existing deep learning methods, such as Faster R-CNN, YOLO series, SSD, RetinaNet and other machine image processing technologies, the scarce features of small target groups are more likely to be lost and failed in the transmission of multi-layer learning networks, which is the root cause of the current bottleneck of high-precision recognition of small targets. Not only that, this type of method also requires the establishment of a large-scale learning network model, which incurs high computing power and manpower costs. It is difficult to meet the accuracy requirements of detection, and it is also difficult to meet the lightweight requirements of edge deployment.

[0003] The current attention mechanism empowers machine learning methods, enhances the ability to capture key features of the target, and alleviates the problem of target positioning and classification. However, this type of method ignores the ability of the target position and the correlation between the target and the surrounding objects to participate in auxiliary decision-making. At the same time, there is a lack of lightweight fusion means to extract feature information between different levels. Therefore, both the high-computing power solution and the lightweight solution lack the assistance of location information and lightweight feature fusion, resulting in the efficiency and accuracy of dense small target detection not being significantly improved. Another way to improve detection accuracy is to fully integrate the task alignment method of classification and positioning information. However, when the target classification and positioning information are inconsistent, there is a certain loss in its weighted assignment, especially in the case of dense target recognition. This weighted assignment is random and volatile, so the effect of improving recognition accuracy is limited. Therefore, how to improve the accuracy of small target detection, expand the scope of application and lightweight model has become a technical problem that needs to be solved. Summary of the invention

[0004] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a high-confidence intelligent detection method, device and medium for clustered dense small targets. By constructing an HLS-YOLO network and using the channel pruning method to improve the efficiency of the improved model, the optimal and efficient HLS-YOLO-P model is obtained, and dense and highly similar small targets are efficiently detected, while the accuracy is improved and the lightweight of the model is ensured.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] According to one aspect of the present invention, a high-confidence intelligent detection method for densely clustered small targets is provided, comprising the following steps:

[0007] S1. Collecting image data through electric power visual inspection equipment and performing preprocessing, wherein the image data includes a target scene;

[0008] S2. Input the preprocessed data into the HLS-YOLO-P network, and perform the small target detection task through the shared parameter task alignment detection head STADH. The output results include target category, location and confidence.

[0009] S3, post-processing the output result, including adjusting the intersection-and-union ratio and the confidence threshold;

[0010] The candidate frames whose intersection-over-union ratio exceeds a preset threshold are regarded as redundant frames and removed, and the frames with the highest confidence and no overlap are retained;

[0011] Set the confidence threshold according to the inspection task requirements and retain the detection boxes that exceed the confidence threshold;

[0012] S4. Output the post-processed detection results to the system interface, display the location and category information of the small target, generate an inspection report according to the needs of power inspection, and trigger the corresponding automatic alarm system according to the abnormal status.

[0013] Furthermore, the preprocessing of S1 includes image size adjustment, brightness contrast adjustment and noise filtering.

[0014] Furthermore, the HLS-YOLO-P network of S2 includes a wavelet transform-based HWD downsampling module for retaining small target detail features, including converting features through two-dimensional discrete Haar wavelet transform and reducing the image spatial resolution; converting the number of channels through the CBS module to match the subsequent layers and filtering redundant information.

[0015] The HLS-YOLO-P network also includes a global feature extraction module C2f-LSKA, which replaces the Darknet Bottleneck structure in the original YOLOv8 network C2f module with the LSKA Block structure, and divides the data into a direct output branch and a branch with increased nonlinear ability and representation ability.

[0016] The LSKA Block structure includes a first 1×1 convolution, an LSKA attention mechanism and a second 1×1 convolution, wherein the first 1×1 convolution provides a nonlinear transformation through a SiLU activation function.

[0017] The HLS-YOLO-P network also includes a task alignment detection head STADH with shared parameters. Specifically, it includes: using a series of convolutional layers with residual connections to fuse multiple layers of feature maps to perform deep feature interaction between positioning and classification tasks; dividing the joint features into positioning and classification branches, and learning weight coefficients specific to the current task through Task Block respectively; outputting the positioning prediction result in the positioning branch; in the classification branch, dynamic feature selection is performed through interactive features to output the classification prediction result.

[0018] The positioning prediction result expression is:

[0019] P align =conv1(Z task ),

[0020] Among them, conv1 is a standard 1×1 convolution, Z task is the intermediate feature map;

[0021] The classification prediction result expression is:

[0022] B align =conv1(M·Z task ),

[0023] M = σ(conv2(δ(conv1(X inter )))),

[0024] Among them, M is the interactive feature weight, conv2 is 3×3 convolution, X inter For N The result of splicing, represents the k-th layer feature map, conv1 is a standard 1×1 convolution, Z task is the intermediate feature map, δ is the ReLU activation function, and σ is the SiLU activation function.

[0025] When constructing the HLS-YOLO-P network, model compression is achieved through channel pruning.

[0026] According to a second aspect of the present invention, there is provided an electronic device, comprising a memory and a processor, wherein a computer program is stored in the memory, and the method described above is implemented when the processor executes the program.

[0027] According to a third aspect of the present invention, there is provided a computer-readable storage medium having a computer program stored thereon, wherein the program implements the method described above when executed by a processor.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] (1) Improving the detection accuracy of small targets: The HWD downsampling module based on wavelet transform can effectively retain the detailed features of small targets, and by integrating the LSKA attention mechanism into the C2f structure, the model can better extract global features and high-level semantic information, thereby improving the detection accuracy and precision of small targets.

[0030] (2) Achieve model lightweighting: By using channel pruning to lightweight the improved model, the model size is reduced, the inference speed is increased, the real-time performance of the model is improved, and the deployment requirements on edge devices are met.

[0031] (3) Improved applicability: In highly complex and dense target scenes, it can accurately locate and identify target classifications, and has good performance in power visual inspection and three-dimensional reconstruction of fine small target detection tasks in different scenes. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 A flowchart of a high-confidence intelligent detection method for densely clustered small objects;

[0033] Figure 2 This is the flow chart for building the HLS-YOLO-P network model;

[0034] Figure 3 The structure diagram of the task alignment detection head STADH;

[0035] Figure 4 This is the architecture diagram of the efficient detection model HLS-YOLO-P;

[0036] Figure 5 This is the detection result diagram of the original YOLOv8n model;

[0037] Figure 6 This is the detection result diagram of the HLS-YOLO model. DETAILED DESCRIPTION

[0038] The following will be combined with the drawings in 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 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 creative work should fall within the scope of protection of the present invention.

[0039] This embodiment provides a high-confidence intelligent detection method for clustered dense small targets, and this method can be applied to the detection of surface status of power inspection switch cabinets. There are many types of equipment in the surface status detection system of power inspection switch cabinets, and the identification targets include various indicator lights in instruments such as digital protection devices, temperature and humidity controllers, and power display instruments, as well as on the surface of power switch cabinets. Among them, the status indicator lights in the instruments are dense small targets. The above types of targets can fully reflect the current operating status of the inspected line. Figure 1 As shown, a high-confidence intelligent detection method for densely clustered small targets includes the following steps:

[0040] S1. Collecting image data through electric power visual inspection equipment and performing preprocessing, wherein the image data includes a target scene;

[0041] S2. Input the preprocessed data into the HLS-YOLO-P network, and perform the small target detection task through the shared parameter task alignment detection head STADH. The output results include target category, location and confidence.

[0042] S3, post-processing the output result, including adjusting the intersection-and-union ratio and the confidence threshold;

[0043] The candidate frames whose intersection-over-union ratio exceeds a preset threshold are regarded as redundant frames and removed, and the frames with the highest confidence and no overlap are retained;

[0044] Set the confidence threshold according to the inspection task requirements and retain the detection boxes that exceed the confidence threshold;

[0045] S4. Output the screened detection results to the system interface, display the location and category information of the small target, generate an inspection report according to the needs of power inspection, and trigger an alarm according to abnormal conditions.

[0046] Data preprocessing of S1 includes image resizing, brightness contrast adjustment, and noise filtering. Perform preliminary screening of images to remove blurred, incomplete, or overly noisy images; perform uniform resizing and normalization of images to achieve consistent image size in the data set; use the imgaug open source image enhancement library to scale, rotate, change brightness, and add Gaussian noise to images to simulate inspection scenes under different environments and lighting conditions, and improve the generalization ability of subsequent training models; use the LabelImg tool to annotate targets, add labels to each annotation box, and accurately reflect the category and status of the target. Categories include overcurrent I protection input, overcurrent I protection start, overload protection input, overload protection start, temperature and humidity controller heating I start, reclosing protection start, and other power switch cabinet surface status indicators that reflect the line status. The technical solution training model of this embodiment randomly divides the expanded and labeled images into training set, verification set, and test set according to the ratio of 7:2:1.

[0047] like Figure 2 As shown in the figure, the HLS-YOLO network is constructed, and the channel pruning method is used to improve the efficiency of the improved model to obtain the optimal and efficient HLS-YOLO-P model.

[0048] The HLS-YOLO-P network includes a HWD downsampling module based on wavelet transform to preserve the details of small targets. Specifically, the size of the input feature map x, that is, the length, width and number of channels of the feature map, is H×W×C. The two-dimensional discrete Haar wavelet transform is used to transform the features and reduce the image spatial resolution. The feature map size is halved to The number of channels is increased to 4C. The generated components pass through the CBS module, i.e., standard 1×1 convolution, batch normalization, and SiLU activation function, to convert the number of channels to C′ to match the subsequent layers and filter out redundant information, so that the subsequent layers can efficiently learn representative features.

[0049] In the backbone network, HWD can retain detailed information such as the shape and texture of the indicator light as much as possible while downsampling, taking into account both computational efficiency and model performance, ensuring efficient feature extraction of small-scale targets in power inspection without significantly increasing the computational burden.

[0050] The HLS-YOLO-P network also includes a global feature extraction module C2f-LSKA, which replaces the Darknet Bottleneck structure in the original YOLOv8 network C2f module with the LSKA Block structure. The basic modules in the LSKA Block are composed of 1×1 convolution, LSKA attention and 1×1 convolution, in which the SiLU activation function is used in the first convolution layer to provide a nonlinear transformation. The processed feature map is residually connected to the original input feature map to maintain the consistency and stability of the features. Through the Split operation in the C2f-LSKA structure, the input data is divided into two branches for processing, one of which is directly output, and the other branch is processed by multiple LSKA Block modules to increase the nonlinear ability and representation ability of the network, thereby improving the network's modeling ability for complex data. In this embodiment, in order to balance the learning effect of long-distance dependencies and the calculation speed, the LSKA convolution kernel is 7.

[0051] like Figure 3 As shown in Figure 1, the HLS-YOLO-P network also includes a task alignment detection head STADH with shared parameters. The concatenated convolutional layers with residual connections are used to fuse multiple layers of feature maps to perform deep feature interaction between positioning and classification tasks, while capturing both detail information and global information to improve spatial resolution. The expression is:

[0052]

[0053] in, is the intermediate feature map of the kth layer, and CBS1 is a convolution block including standard 3×3 convolution, batch normalization, and SiLU activation function.

[0054] The joint features are divided into positioning and classification branches, and the weight coefficients specific to the current task are learned through Task Block. Global information is extracted from each feature map through global average pooling to reduce the number of parameters and calculations of the model while retaining key feature information. Then, through convolution and activation, the weight coefficients suitable for the current task are learned to dynamically adjust the importance of the feature map. The expression is:

[0055]

[0056] w = σ(conv1(δ(conv1(x inter )))),

[0057] Z task =CBS2(X task ),

[0058] in, is the k-th layer feature map, x inter For After splicing, X task As a result of global average pooling, conv1 is a standard 1×1 convolution, δ is the ReLU activation function, σ is the SiLU activation function, w is the vector used to generate the overall weight, w k is the weight of a specific layer, which is used to adjust the feature map of the kth layer, Z task is the intermediate feature map, and CBS2 is a convolution block including standard 1×1 convolution, batch normalization, and SiLU activation function.

[0059] The positioning prediction results are output in the positioning branch; in the classification branch, dynamic feature selection is performed through interactive features to output the classification prediction results.

[0060] The positioning prediction result expression is:

[0061] P align =conv1(Z task ),

[0062] Among them, conv1 is a standard 1×1 convolution, Z task is the intermediate feature map;

[0063] The classification prediction result expression is:

[0064] B align =conv1(M·Z task ),

[0065] M = σ(conv2(δ(conv1(X inter )))),

[0066] Among them, M is the interactive feature weight, conv2 is 3×3 convolution, X inter For N The result of splicing, represents the k-th layer feature map, conv1 is a standard 1×1 convolution, Z task is the intermediate feature map, δ is the ReLU activation function, and σ is the SiLU activation function.

[0067] Therefore, the steps for improving the original YOLOv8 network model to HLS-YOLO in this embodiment are as follows: introduce the HWD module and replace the first and second convolutional downsampling structures of the original backbone network structure, with the feature map size and channel output being 160×160×32 and 80×80×64 respectively; integrate the LSKA attention mechanism into the C2f structure and replace the last C2f module of the backbone network; replace the original detection head of the network with the STADH parameter sharing and task alignment detection head.

[0068] For the improved network HLS-YOLO, the dataset obtained from mgaug open source images is used for model training, and an improved detection model for dense and highly similar small targets of power switch cabinets is obtained. Regularization is used to achieve the sparsification of the improved network HLS-YOLO. Specifically, the L1 regularization term is used to make the value of the scaling factor in the BN layer tend to 0 to identify unimportant channels. The loss function expression of sparse training is:

[0069]

[0070] Among them, l(f(x,W),y) is the loss function of the network before network sparseness, is the regularization term, λ is the weight, Γ is the set of scaling factors for all BN layers in the improved YOLOv8 network, and g(·) is the L1 regularization.

[0071] Perform channel pruning on the sparsified network. In the CBS structure of the HLS-YOLO algorithm, the scaling factor in the batch normalization layer directly affects the activation scale of each convolution channel. A smaller scaling factor will result in smaller batch normalized output and output after the activation function, that is, the channel corresponding to the scaling factor contributes less to the network. After sparse training, the scaling factor of the improved model will be concentrated near 0, thereby filtering out channels that contribute less to the network. During pruning, the channels with lower contributions and their corresponding convolution kernels are pruned to achieve more efficient model compression.

[0072] Use the fine-tuning operation to fine-tune the pruned network model to balance the detection speed and accuracy. Specifically, the recognition effect calculation expression is:

[0073]

[0074]

[0075]

[0076]

[0077] Among them, N TP is the total number of correctly classified positive samples, N FP is the number of negative samples misclassified as positive, N FN is the number of unrecognized positive samples, N TP +N FP is the total number of samples detected as positive, N TP +N FN is the total number of actual positive samples. Construct a PR curve, and the value of each point on the PR curve is P(r). The average precision AP can be obtained by the integration method, where AP 50is the average accuracy of detection under the threshold of 0.5 when the intersection of predicted box and true box is equal to 0.5. n is the number of types of targets to be detected. In addition, the amount of calculation, the number of parameters and the processing speed of the model are also considered comprehensively, that is, the number of frames processed per second. The higher the number of frames processed per second, the faster the detection speed of the model.

[0078] In this embodiment, by adjusting the speedup ratio, that is, the ratio of the amount of computation after pruning to the original amount of computation, different degrees of pruning of the network are achieved. The results of the pruning experiment are shown in Table 1.

[0079] Table 1 Pruning experimental results with different speedup ratios

[0080]

[0081] As the acceleration ratio increases, the channels with smaller contributions in the network and their corresponding convolution kernels will be gradually pruned away to obtain a more streamlined detection network. As can be seen from Table 1, when the acceleration ratio increases, the number of network parameters continues to decrease, and the model size continues to decrease. When the acceleration ratio is within 2.2, the mAP50 of the model decreases slightly; when the acceleration ratio reaches 2.4 and 2.6, the mAP50 decreases by 1.2% and 2.1% respectively, and the inference speed is not significantly improved. When the acceleration ratio is 2.2, the number of frames processed per second increases by 17.1 frames, while the mAP50 only decreases by 0.5%. Therefore, this embodiment ultimately selects an acceleration ratio of 2.2 for pruning operations, and fine-tunes the pruned model to obtain the final efficient detection model HLS-YOLO-P. The architecture diagram of the efficient detection model HLS-YOLO-P is shown in the figure below. Figure 4 shown.

[0082] like Figure 5 and Figure 6 The figure shows the detection results before and after the model improvement. Among them, Cls is classification error, Miss is missed detection error, Loc is positioning error, Bkg is background error; Both is classification and positioning error, Dupe; repeated detection error, FN is false negative error, FP is false positive error. When using the original YOLOv8n model for power switch cabinet scene indicator light recognition, the main factors limiting the detection effect are missed detection and classification errors. When HLS-YOLO is used for detection, the classification error is greatly reduced by 2.11%, and the missed detection error is reduced from 3.98% to 2.87%, indicating that the improved HLS-YOLO network has significantly improved its ability to detect small targets.

[0083] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the described module can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0084] The electronic device of the present invention includes a central processing unit (CPU), which can perform various appropriate actions and processes according to computer program instructions stored in a read-only memory (ROM) or loaded from a storage unit into a random access memory (RAM). In the RAM, various programs and data required for device operation can also be stored. The CPU, ROM and RAM are connected to each other via a bus. An input / output (I / O) interface is also connected to the bus.

[0085] A plurality of components in the device are connected to the I / O interface, including: an input unit, such as a keyboard, a mouse, etc.; an output unit, such as various types of displays, speakers, etc.; a storage unit, such as a disk, an optical disk, etc.; and a communication unit, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit allows the device to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks. The processing unit performs the various methods and processes described above, such as the method of the present invention. For example, in some embodiments, the method of the present invention may be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as a storage unit. In some embodiments, part or all of the computer program may be loaded and / or installed on the device via a ROM and / or a communication unit. When the computer program is loaded into the RAM and executed by the CPU, one or more steps of the method of the present invention described above may be performed. Alternatively, in other embodiments, the CPU may be configured to perform the method of the present invention in any other appropriate manner (e.g., by means of firmware).

[0086] The functions described above herein may be performed at least in part by one or more hardware logic components. For example, without limitation, exemplary types of hardware logic components that may be used include: field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), complex programmable logic devices (CPLDs), and the like.

[0087] The program code for implementing the method of the present invention can be written in any combination of one or more programming languages. These program codes can be provided to a processor or controller of a general-purpose computer, a special-purpose computer or other programmable data processing device, so that the program code, when executed by the processor or controller, enables the functions / operations specified in the flow chart and / or block diagram to be implemented. The program code can be executed entirely on the machine, partially on the machine, partially on the machine as a stand-alone software package and partially on a remote machine, or entirely on a remote machine or server.

[0088] In the context of the present invention, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or equipment. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0089] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A high-confidence intelligent detection method for densely clustered small targets, characterized in that: The following steps are involved: S1. Collecting image data through electric power visual inspection equipment and performing preprocessing, wherein the image data includes a target scene; S2. Input the preprocessed data into the HLS-YOLO-P network, and perform the small target detection task through the shared parameter task alignment detection head STADH. The output results include target category, location and confidence. S3, post-processing the output result, including adjusting the intersection-and-union ratio and the confidence threshold; The candidate frames whose intersection-over-union ratio exceeds a preset threshold are regarded as redundant frames and removed, and the frames with the highest confidence and no overlap are retained; Set the confidence threshold according to the inspection task requirements and retain the detection boxes that exceed the confidence threshold; S4. Output the post-processed detection results to the system interface, display the location and category information of the small target, generate an inspection report according to the needs of power inspection, and trigger an alarm according to abnormal conditions.

2. The high-confidence intelligent detection method for densely clustered small targets according to claim 1 is characterized in that: The preprocessing of S1 includes image size adjustment, brightness and contrast adjustment, and noise filtering.

3. The high-confidence intelligent detection method for densely clustered small targets according to claim 1 is characterized in that: The HLS-YOLO-P network of S2 includes a HWD downsampling module based on wavelet transform, which is used to retain the detailed features of small targets, including converting features through two-dimensional discrete Haar wavelet transform and reducing the image spatial resolution; converting the number of channels through the CBS module to match the subsequent layers and filtering redundant information.

4. The high-confidence intelligent detection method for densely clustered small targets according to claim 3 is characterized in that: The HLS-YOLO-P network also includes a global feature extraction module C2f-LSKA, which replaces the Darknet Bottleneck structure in the original YOLOv8 network C2f module with the LSKA Block structure, and divides the data into a direct output branch and a branch with increased nonlinear ability and representation ability.

5. The high-confidence intelligent detection method for densely clustered small targets according to claim 4 is characterized in that: The LSKA Block structure includes a first 1×1 convolution, an LSKA attention mechanism and a second 1×1 convolution, wherein the first 1×1 convolution provides a nonlinear transformation through a SiLU activation function.

6. The high-confidence intelligent detection method for densely clustered small targets according to claim 4 is characterized in that: The HLS-YOLO-P network also includes a task alignment detection head STADH with shared parameters, Specifically, it includes: using serial convolutional layers with residual connections to fuse multiple layers of feature maps to perform deep feature interaction between positioning and classification tasks; dividing the joint features into positioning and classification branches, and learning weight coefficients specific to the current task through Task Block respectively; outputting the positioning prediction results in the positioning branch; in the classification branch, dynamic feature selection is performed through interactive features to output the classification prediction results.

7. A high-confidence intelligent detection method for densely clustered small targets according to claim 6, characterized in that: The positioning prediction result expression is: P align =conv1(Z task ), Among them, conv1 is a standard 1×1 convolution, Z task is the intermediate feature map; The classification prediction result expression is: B align =conv1(M·Z task ), M=σ(conv2(δ(conv1(X inter )))), Among them, M is the interactive feature weight, conv2 is a 3×3 convolution, and x inter For After splicing, X task As the result of global average pooling, represents the k-th layer feature map, conv1 is a standard 1×1 convolution, Z task is the intermediate feature map, δ is the ReLU activation function, and σ is the SiLU activation function.

8. The high-confidence intelligent detection method for densely clustered small targets according to claim 1, characterized in that: When constructing the HLS-YOLO-P network, model compression is achieved through channel pruning.

9. An electronic device comprising a memory and a processor, wherein a computer program is stored in the memory, wherein: When the processor executes the program, the method according to any one of claims 1 to 8 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 8 is implemented.

Images