Small target detection method and device for power scene, equipment and medium

Abstract

The invention provides a small target detection method and device for a power scene, equipment and a medium, and relates to the technical field of image processing, in particular to the field of target detection. The method comprises the following steps: inputting a to-be-detected image of a power scene into a first receptive field enhancement module group after passing through a first convolutional layer to obtain a first enhanced feature map; inputting the first enhanced feature map into a second receptive field enhancement module group to obtain a second enhanced feature map; inputting the second enhanced feature map into a spatial pyramid pooling layer after passing through a second convolutional layer to obtain a pooling feature map; inputting the pooling feature map into a first multi-scale feature fusion module to obtain a first fusion feature map; and inputting the first fusion feature map into a first detection head to obtain a first target detection result, the first detection head being used for detecting whether the to-be-detected image includes a small target of a preset first scale. The small target detection accuracy of the electric power scene can be improved.

Description

Technical Field

[0001] The present application relates to the field of image processing technology, in particular to the field of target detection, and specifically to a method, device, equipment and medium for detecting small targets in power scenarios. Background Art

[0002] In power scenarios, safety monitoring is a key factor in ensuring the safety of workers and equipment. Traditional manual inspection methods have problems such as high labor costs, low monitoring efficiency, and difficulty in covering a large area.

[0003] At present, deep learning has made remarkable achievements in the field of target detection and can provide more accurate and efficient solutions. However, for small target objects in power scenes, traditional deep learning methods still have some challenges. Small target objects often have low pixel resolution and rich background interference, which puts higher requirements on the accuracy and robustness of small target object detection in power scenes. Summary of the invention

[0004] The present application provides a small target detection method, device, equipment and medium in an electric power scene, so as to solve the problem of low accuracy of small target detection in an electric power scene in the prior art.

[0005] In a first aspect, the present application provides a small target detection method in an electric power scenario, comprising:

[0006] Input the image to be detected of the power scene into the first receptive field enhancement module group after passing through the first convolution layer to obtain a first enhanced feature map;

[0007] Inputting the first enhanced feature map into a second receptive field enhancement module group to obtain a second enhanced feature map;

[0008] Input the second enhanced feature map into the spatial pyramid pooling layer after passing through the second convolution layer to obtain a pooled feature map;

[0009] Inputting the pooled feature map into a first multi-scale feature fusion module to obtain a first fused feature map;

[0010] The first fusion feature map is input into a first detection head to obtain a first target detection result, wherein the first detection head is used to detect whether the image to be detected includes a small target of a preset first scale.

[0011] In a second aspect, the present application provides a small target detection device for an electric power scenario, comprising:

[0012] A first enhancement module, used for inputting the image to be detected of the power scene into the first receptive field enhancement module group after passing through the first convolution layer, to obtain a first enhanced feature map;

[0013] A second enhancement module, configured to input the first enhanced feature map into a second receptive field enhancement module group to obtain a second enhanced feature map;

[0014] A pooling module, used for inputting the second enhanced feature map into a spatial pyramid pooling layer after passing through a second convolutional layer to obtain a pooled feature map;

[0015] A first fusion module, used for inputting the pooled feature map into a first multi-scale feature fusion module to obtain a first fused feature map;

[0016] The first detection module is used to input the first fusion feature map into a first detection head to obtain a first target detection result, wherein the first detection head is used to detect whether the image to be detected includes a small target of a preset first scale.

[0017] In a third aspect, the present application also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, a small target detection method for power scenarios as described in any one of the embodiments of the present application is implemented.

[0018] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a small target detection method for an electric power scenario as described in any of the embodiments of the present application.

[0019] In a fifth aspect, the present application also provides a computer program product, including a computer program, which, when executed by a processor, implements a small target detection method for an electric power scenario as described in any one of the embodiments of the present application.

[0020] The small target detection method, device, equipment and medium of the power scene provided by the present application, the image to be detected of the power scene is sequentially input into two receptive field enhancement module groups after the convolution layer, and the receptive field is enhanced to better perceive the image information of a larger range and capture features of different scales. Then, the features output by the receptive field enhancement module group are respectively subjected to multi-scale feature fusion after the convolution layer and the pyramid pooling layer, and finally a fused feature map is obtained, and the detection head performs small target detection according to the fused feature map. Thus, the network's perception ability of small targets can be improved, so that the network can better capture small targets in power scenes and improve the accuracy of small target detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0022] Figure 1 A flowchart of a small target detection method in an electric power scenario provided in an embodiment of the present application;

[0023] Figure 2 A schematic diagram of a small target detection network structure provided in an embodiment of the present application;

[0024] Figure 3 A flowchart of the receptive field enhancement process in the small target detection method provided in an embodiment of the present application;

[0025] Figure 4 A schematic diagram of a receptive field enhancement module provided in an embodiment of the present application;

[0026] Figure 5 A schematic diagram of the structure of a small target detection device for an electric power scenario provided in an embodiment of the present application;

[0027] Figure 6 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application.

[0028] The above drawings have shown clear embodiments of the present application, which will be described in more detail later. These drawings and text descriptions are not intended to limit the scope of the present application in any way, but to illustrate the concept of the present application to those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0029] The present application will be further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the present application, rather than to limit the present application. It should also be noted that, for ease of description, only the parts related to the present application, rather than all structures, are shown in the accompanying drawings.

[0030] It should be noted that similar numbers and letters represent similar items in the following figures. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. At the same time, in the description of this application, the terms "first", "second", etc. are only used to distinguish the description and cannot be understood as indicating or implying relative importance. The acquisition, storage, use, processing, etc. of data in the technical solution of this application comply with the relevant provisions of national laws and regulations. It should be noted that in the embodiments of this application, certain software, components, models and other existing solutions in the industry may be mentioned, and they should be regarded as exemplary. The purpose is only to illustrate the feasibility of the implementation of the technical solution of this application, but it does not mean that the applicant has or must use the relevant content of the solution.

[0031] Figure 1 Flow chart of a small target detection method for power scene provided by an embodiment of the present application. This embodiment is applicable to the case of detecting small targets in power scenes. The method can be performed by a small target detection device for power scenes. The device can be implemented in software and / or hardware, preferably configured in an electronic device, such as a computer device or a server. Figure 1 As shown, the method specifically includes:

[0032] S101. Input the image to be detected of the power scene into the first receptive field enhancement module group after passing through the first convolution layer to obtain a first enhanced feature map.

[0033] S102: Input the first enhanced feature map into the second receptive field enhancement module group to obtain a second enhanced feature map.

[0034] S103, input the second enhanced feature map into the spatial pyramid pooling layer after passing through the second convolution layer to obtain a pooled feature map.

[0035] S104: Input the pooled feature map into a first multi-scale feature fusion module to obtain a first fused feature map.

[0036] S105. Input the first fusion feature map into the first detection head to obtain a first target detection result.

[0037] Specifically, small targets in power scenarios include objects of different sizes such as electrician testers, goggles or helmets used by workers. Whether these small targets can be used in a planned manner is very critical. Therefore, accurate detection of small target objects in power scenarios plays a key role in the safety of workers and equipment. In this application, a small target detection network based on receptive field enhancement is established, which can improve the accuracy and stability of small target detection in power scenarios.

[0038] In this embodiment, the small target detection network based on receptive field enhancement includes a first convolutional layer, a first receptive field enhancement module group, a second receptive field enhancement module group, a second convolutional layer, a first multi-scale feature fusion module and a first detection head.

[0039] First, the image to be detected of the power scene is input into the first convolution layer, and the number of channels of the feature map of the image to be detected is adjusted through the first convolution layer to enrich the expression ability of the features so that the subsequent modules can capture more visual information. Then the output result of the first convolution layer is input into the first receptive field enhancement module group to obtain the first enhanced feature map. The first receptive field enhancement module group can be composed of one or more receptive field enhancement modules, which are used to perceive a larger range of image information and capture features of different scales at the same time, thereby improving the adaptability and detection performance of the network to complex scenes and better detecting small target objects. Then, the first enhanced feature map is input into the second receptive field enhancement module group to obtain the second enhanced feature map. The second receptive field enhancement module group can be the same as the first receptive field enhancement module group or different. The receptive field of the first enhanced feature map is further enhanced by the second receptive field enhancement module group. Then, the second enhanced feature map is input into the spatial pyramid pooling layer after the second convolution layer to obtain the pooled feature map, and the pooled feature map is input into the first multi-scale feature fusion module to obtain the first fused feature map. Among them, the second convolutional layer is used to further fuse and enhance the multi-scale features in the second enhanced feature map of the input, deepen the feature extraction, and adjust the size of the feature map. Through the spatial pyramid pooling layer and the first multi-scale feature fusion module, the input features can be further reasonably fused, so that the network can more comprehensively perceive the small target objects in the power operation scene, and understand the image information from different scales, and obtain a first fused feature map containing multi-scale and rich information. Finally, the first fused feature map is input into the first detection head to obtain the first target detection result. Among them, the first detection head is used to detect whether the image to be detected includes a small target of a preset first scale. The first scale can be predetermined based on the scale division of the detection object.

[0040] The small target detection method, device, equipment and medium of the power scene provided by the present application, the image to be detected of the power scene is sequentially input into two receptive field enhancement module groups after the convolution layer, and the receptive field is enhanced to better perceive the image information of a larger range and capture features of different scales. Then, the features output by the receptive field enhancement module group are respectively subjected to multi-scale feature fusion after the convolution layer and the pyramid pooling layer, and finally a fused feature map is obtained, and the detection head performs small target detection according to the fused feature map. Thus, the network's perception ability of small targets can be improved, so that the network can better capture small targets in power scenes and improve the accuracy of small target detection.

[0041] On the basis of the above embodiments, the small target detection method of the power scene of the present application also includes: inputting the pooled feature map into the third convolution layer, and performing feature splicing on the output result of the third convolution layer and the second enhanced feature map to obtain a first spliced ​​feature map; inputting the first spliced ​​feature into the second multi-scale feature fusion module to obtain a second fused feature map; and inputting the second fused feature map into the second detection head to obtain a second target detection result. The second detection head is used to detect whether the image to be detected includes a small target of a preset second scale.

[0042] Among them, the third convolutional layer can be the same as the first convolutional layer or the second convolutional layer, or it can be different, and is used to adjust the number of channels of the input feature map. The second scale is different from the first scale. In this embodiment, the output result of the third convolutional layer is firstly feature-spliced ​​with the second enhanced feature map to obtain a first spliced ​​feature map, and then the first spliced ​​feature is input into the second multi-scale feature fusion module to obtain a second fused feature map. Compared with the first fused feature map, the second fused feature map has more scales, richer image features and information, so that the small targets of the second scale can be better detected by the second detection head.

[0043] Further, based on the above embodiments, the small target detection method for the power scene of the present application also includes: inputting the second fused feature map into the fourth convolution layer, and performing feature splicing on the output result of the fourth convolution layer and the first enhanced feature map to obtain a second spliced ​​feature map; inputting the second spliced ​​feature into the third multi-scale feature fusion module to obtain a third fused feature map; and inputting the third fused feature map into the third detection head to obtain a third target detection result. The third detection head is used to detect whether the image to be detected includes a small target of a preset third scale.

[0044] The third scale is different from the second scale and the first scale. In order to more accurately detect small targets of the third scale, this embodiment further performs feature splicing and fusion based on the second fusion feature, thereby improving the accuracy of detecting small targets of the third scale.

[0045] Figure 2A schematic diagram of the small target detection network structure provided for an embodiment of the present application. The small target detection network includes three parts, namely a trunk receptive field enhancement network 1, a neck feature fusion network 2, and a head detection network 3. Among them, the trunk receptive field enhancement network 1 is the core part of the power scene small target detection network based on receptive field enhancement, and is composed of a stacked combination of a convolution module and a receptive field enhancement module. Specifically includes convolution 11, receptive field enhancement modules 12-15 and convolution 16. The neck feature fusion network 2 is located between the trunk receptive field enhancement network 1 and the head detection network 3, and is composed of a spatial pyramid pooling module and multiple multi-scale feature fusion modules, a convolution module, and a feature splicing module, which is used to further optimize and refine the features extracted by the trunk receptive field enhancement network. Specifically includes spatial pyramid pooling 21, multi-scale feature fusion module 22, convolution 23, feature splicing 24, multi-scale feature fusion module 25, convolution 26, feature splicing 27 and multi-scale feature fusion module 28. The head detection network 3 is the last part of the small target detection network for power scenes based on receptive field enhancement, which is used to generate a target detection frame, provide classification information and confidence information of the small target, and determine the small target detection result. Specifically, it includes a small detection head 31, a medium detection head 32, and a large detection head 33. Different detection heads are used to detect small targets of different scales in the image. In one embodiment, different detection heads have different sizes of detection frames for small target detection.

[0046] Specifically, the image to be detected is input to the convolution 11 to adjust the number of channels of the feature map, and its output results are sequentially passed through the receptive field enhancement modules 12-15 to obtain a larger range of image information and capture features of different scales in the image. The result is input to the control pyramid pooling 21 after convolution 16, and then the first fused feature map is obtained by the multi-scale feature fusion module 22. The small detection head detects small targets of the first scale based on the first fused feature map. Furthermore, for small targets of the second scale, the output result of the spatial pyramid pooling 21 is convolved 23, and then the output result of the receptive field enhancement module 15 is feature spliced ​​by feature splicing 24, and then fused by the multi-scale feature fusion module 25 to obtain a second fused feature map, and the middle detection head 32 detects small targets of the second scale based on the second fused feature map. For small targets of the third scale, the output of the multi-scale feature fusion module 25 is further processed by convolution 26, and then feature stitching 27 is used with the output result of the receptive field enhancement module 13, and then the third fusion feature map is obtained by the multi-scale feature fusion module 28. The large detection head 33 detects small targets of the third scale based on the third fusion feature map.

[0047] Usually, small target objects have lower resolution and rich background interference, so it is difficult to detect. However, the embodiment of the present application further divides the small targets of the power scene into three types according to different scales, and establishes a corresponding detection head, which enhances the receptive field by stacking different numbers of receptive field enhancement modules, and also splices and fuses different feature maps to obtain fused feature maps of different scales. Through the processing of these convolution modules and feature splicing modules, the fused features are further strengthened and enriched to obtain more advanced and abstract semantic information, thereby improving the detection effect of small targets of different scales and meeting the detection needs of small targets of different scales. Therefore, the present application can be applied to more complex power scenes, and can simultaneously accurately detect small targets of different scales. When the same image to be detected contains multiple small targets of different scales, the present application can be based on the splicing results of different numbers of stacked receptive field enhancement modules and different scale features, for the respective characteristics of small targets of different scales, to achieve accurate detection.

[0048] Figure 3 A flowchart of the receptive field enhancement process in the small target detection method provided in the embodiment of the present application. In this embodiment, the input of any receptive field enhancement module is used as the current feature map, and the process of the receptive field enhancement module outputting the current enhanced feature map according to the current feature map includes:

[0049] S301. Input the current feature map into the fifth convolutional layer to obtain a first convolutional feature map, and input the first convolutional feature map into the sixth convolutional layer to obtain a second convolutional feature map.

[0050] S302: perform feature concatenation on the first convolution feature map and the second convolution feature map to obtain a third concatenated feature map.

[0051] S303, sequentially pass the third concatenated feature map through the maximum pooling layer, the attention enhancement layer, the seventh convolutional layer and the eighth convolutional layer to obtain a third convolutional feature map.

[0052] S304: perform feature splicing on the first convolution feature map, the second convolution feature map, and the third convolution feature map to obtain a fourth spliced ​​feature map.

[0053] S305: Perform feature splicing on the fourth spliced ​​feature map and the current feature map to obtain a current enhanced feature map.

[0054] Among them, the feature concatenation method includes: adding feature elements, multiplying feature elements, or adding feature elements first and then multiplying them.

[0055] Figure 4 Schematic diagram of the receptive field enhancement module provided in the embodiment of the present application. Convolution 1 and convolution 2 correspond to the fifth convolution layer and the sixth convolution layer respectively. Convolution 3 and convolution 4 correspond to the seventh convolution layer and the eighth convolution layer respectively.

[0056] Specifically, the current feature map is input into the receptive field enhancement module, and the two convolution modules, convolution 1 and convolution 2, output two feature maps with different numbers of channels and resolutions, namely the first convolution feature map and the second convolution feature map, which enriches the expressive power of the features. Then, the first convolution feature map and the second convolution feature map are spliced ​​by feature splicing 1, for example, by element-wise addition to further fuse them to generate a feature map containing more feature information. Next, through the two modules of maximum pooling and attention enhancement, the network can perceive a wider range of image information and capture features of different scales at the same time, so that the network can better capture the information of small target objects. The two modules of convolution 3 and convolution 4 continue to adjust the channels of the feature map to further optimize the feature representation. The feature maps extracted by convolution 1, convolution 2 and convolution 4 are spliced ​​by feature splicing 2, for example, by further fusing multi-scale features by first adding and then multiplying feature elements. Finally, feature splicing 3 receives the output result of feature splicing 2 and the original current feature map of the input receptive field enhancement module, and further fuses them by multiplying feature elements to enhance these multi-scale features and deepen feature extraction. At the same time, it adjusts the size and number of channels of the feature map to obtain the final output current enhanced feature map.

[0057] Therefore, the receptive field enhancement module in the embodiment of the present application effectively extracts multi-scale features through multi-level feature fusion and receptive field expansion, enhances the model's perception ability of small targets, and improves the performance of small target detection in power scenarios.

[0058] In one embodiment, for the small target detection network with enhanced receptive field constructed in the present application, it is necessary to first collect images and videos of power operation scenarios such as power transmission and power testing, and create an image detection dataset for small target objects in the scene that affect the safety of the operation. These image detection datasets are then used to train the constructed small target detection network to determine the network parameters and obtain the final model.

[0059] First, obtain images and videos taken at the power operation site. For images, factors that affect the safety of power operations can be directly framed as annotation data for training data. For videos, images can be captured frame by frame to obtain images and then annotated. If the scene changes in the video are small, the video can be accelerated and then screenshots and annotations can be made, that is, alternate frame extraction is used to ensure the diversity of data set samples. These image samples and their corresponding annotated data are divided into training sets, validation sets, and test sets according to a certain ratio. And use data enhancement techniques, such as image rotation, flipping, scaling, and brightness adjustment. Random noise can also be added to the image to simulate real scenes under different conditions, such as adding Gaussian noise to simulate scenes with insufficient lighting such as night, simulate extreme weather scenes such as rain or fog, or simulate scenes that interfere with the environment such as trees. The above methods can further expand the data set, increase the number and richness of the data set, and improve the generalization ability of the model.

[0060] Then, the model of the small target detection network is supervised and trained using the data in the training set to determine the network parameters. The trained detection model is deployed to the operation scene monitoring equipment to monitor the operation scene in real time and output the small target detection results. For example, first detect whether the image captured on the spot of the power scene contains a small target and whether the small target is a test pen. If it is a test pen, it can be further determined by whether the test pen is detected a set number of times within a set time period to determine whether the electricity is used in a standardized manner.

[0061] The training process includes three parts: model training, model evaluation and model testing, as follows:

[0062] Model training: Use the training set in the small target image detection dataset to train the model structure of the power scene small target detection network based on receptive field enhancement. Use the defined loss function to measure the difference between the model output and the annotated true label. Then use the back propagation algorithm to update the model parameters to optimize the performance of the model and ensure that small target objects in the power scene can be accurately detected.

[0063] Model evaluation: Use a validation dataset independent of the training dataset to validate the trained model. By comparing the model prediction results with the real labels in the validation dataset, quantify the model's performance indicators such as accuracy, precision, and recall, thereby evaluating its performance on unseen data. Based on the validation results, tune the model parameters to improve the generalization ability of the model and further optimize the detection performance.

[0064] Model testing: Use the test set to test the trained and verified model to evaluate its performance in real scenarios. If it is not ideal, return to the training and verification stage for further tuning until the expected performance level is achieved.

[0065] The loss function is defined as It can be expressed as follows:

[0066]

[0067] Among them, λ represents a learnable hyperparameter used to balance the weight of each part in the loss function. The subscripts loc, cls and obj correspond to the positioning loss, classification loss and confidence loss respectively. p and t represent the predicted value and the labeled true value of the model output respectively. SmoothL1 represents the smoothed L1 loss function, which is used to measure the prediction accuracy of the bounding box position. Reg(θ) represents the regularization term, which is used to control the complexity of the model and prevent overfitting.

[0068] By setting learnable hyperparameters, smooth L1 loss function and regularization terms, the flexibility, robustness and generalization ability of the model are improved. Category weights and adjustment factors are introduced to effectively solve the problems of category imbalance and difficult-to-classify samples, thereby improving the detection accuracy and the ability to detect small targets.

[0069] Furthermore, the classification loss The specific expressions are as follows:

[0070]

[0071] Among them, α represents the category weight, which is used to adjust the importance of each category, and γ represents the adjustment factor, which is used to control the decay rate of the loss. Represents the model's predicted probability vector for each category of the i-th sample, The one-hot encoding vector representing the true category label of the i-th sample, where only the position of the true category is 1 and the rest are 0.

[0072] Furthermore, the confidence loss The specific expressions are as follows:

[0073]

[0074] Among them, β represents the weight of the target category, which is used to adjust the importance of the target category and the non-target category. τ represents the adjustment factor, which is used to control the decay rate of the loss. represents the model's predicted probability that the i-th sample belongs to the target category, Represents the true target category label of the i-th sample, which is 1 if it is the target category, otherwise it is 0.

[0075] In summary, compared with the prior art, the small target detection method for power scenes based on receptive field enhancement of the present application has at least the following beneficial effects:

[0076] (1) Compared with the traditional target detection network, the small target detection network structure in the power scene based on receptive field enhancement of the present application can better capture small targets in the power scene and improve the detection accuracy by enhancing the receptive field.

[0077] (2) The receptive field enhancement module of this application adopts a series of feature extraction and fusion operations, and improves the network's perception of small targets through multi-scale feature enhancement and fusion. This can effectively solve the problems of small targets being easily blocked and having low contrast in power scenes, and improve the robustness of detection.

[0078] (3) This application uses optimized loss functions, including positioning loss, classification loss, and confidence loss. By reasonably setting hyperparameters, introducing category weights and adjustment factors, etc., it can better handle the problems of category imbalance and difficult-to-classify samples, and improve the accuracy and stability of detection.

[0079] Figure 5 A schematic diagram of the structure of a small target detection device for an electric power scenario provided in an embodiment of the present application.

[0080] like Figure 5 As shown, the device 50 comprises:

[0081] The first enhancement module 510 is used to input the image to be detected of the power scene into the first receptive field enhancement module group after passing through the first convolution layer to obtain a first enhanced feature map;

[0082] A second enhancement module 520, configured to input the first enhanced feature map into a second receptive field enhancement module group to obtain a second enhanced feature map;

[0083] A pooling module 530, configured to input the second enhanced feature map into a spatial pyramid pooling layer after passing through a second convolutional layer to obtain a pooled feature map;

[0084] A first fusion module 540, configured to input the pooled feature map into a first multi-scale feature fusion module to obtain a first fused feature map;

[0085] The first detection module 550 is used to input the first fusion feature map into a first detection head to obtain a first target detection result, wherein the first detection head is used to detect whether the image to be detected includes a small target of a preset first scale.

[0086] In some embodiments, the apparatus further comprises:

[0087] A first feature splicing module, used for inputting the pooled feature map into a third convolutional layer, and performing feature splicing on an output result of the third convolutional layer and the second enhanced feature map to obtain a first spliced ​​feature map;

[0088] A first feature fusion module, used for inputting the first splicing feature into a second multi-scale feature fusion module to obtain a second fused feature map;

[0089] The second detection module is used to input the second fusion feature map into a second detection head to obtain a second target detection result, wherein the second detection head is used to detect whether the image to be detected includes a small target of a preset second scale.

[0090] In some embodiments, the apparatus further comprises:

[0091] A second feature splicing module, used for inputting the second fused feature map into a fourth convolutional layer, and performing feature splicing on an output result of the fourth convolutional layer and the first enhanced feature map to obtain a second spliced ​​feature map;

[0092] A second feature fusion module, used for inputting the second spliced ​​features into a third multi-scale feature fusion module to obtain a third fused feature map;

[0093] The third detection module is used to input the third fusion feature map into a third detection head to obtain a third target detection result, wherein the third detection head is used to detect whether the image to be detected includes a small target of a preset third scale.

[0094] In some embodiments, the first receptive field enhancement module group and the second receptive field enhancement module group are each composed of two receptive field enhancement modules, and each receptive field enhancement module is used to expand the perception range of the image and capture features of different scales of the image.

[0095] In some embodiments, the input of any receptive field enhancement module is used as the current feature map, and the process in which the receptive field enhancement module outputs the current enhanced feature map according to the current feature map includes:

[0096] Inputting the current feature map into the fifth convolutional layer to obtain a first convolutional feature map, and inputting the first convolutional feature map into the sixth convolutional layer to obtain a second convolutional feature map;

[0097] Performing feature splicing on the first convolution feature map and the second convolution feature map to obtain a third spliced ​​feature map;

[0098] Passing the third concatenated feature map sequentially through a maximum pooling layer, an attention enhancement layer, a seventh convolutional layer, and an eighth convolutional layer to obtain a third convolutional feature map;

[0099] Performing feature splicing on the first convolution feature map, the second convolution feature map, and the third convolution feature map to obtain a fourth spliced ​​feature map;

[0100] The fourth spliced ​​feature map and the current feature map are feature spliced ​​to obtain the current enhanced feature map.

[0101] In some embodiments, the feature concatenation method includes: adding feature elements, multiplying feature elements, or adding feature elements first and then multiplying them.

[0102] The small target detection device for the power scene provided in the embodiment of the present application can be used to execute the technical solution of the small target detection method for the power scene in the above embodiment. Its implementation principle and technical effect are similar and will not be repeated here.

[0103] It should be noted that it should be understood that the division of the various modules of the above device is only a division of logical functions. In actual implementation, they can be fully or partially integrated into one physical entity, or they can be physically separated. And these modules can all be implemented in the form of software calling through processing elements; they can also be all implemented in the form of hardware; some modules can also be implemented in the form of software calling through processing elements, and some modules can be implemented in the form of hardware. For example, the first enhancement module 510 can be a separately established processing element, or it can be integrated in a chip of the above device. In addition, it can also be stored in the memory of the above device in the form of program code, and called and executed by a processing element of the above device. The function of the first enhancement module 510. The implementation of other modules is similar. In addition, all or part of these modules can be integrated together, or they can be implemented independently. The processing element here can be an integrated circuit with signal processing capabilities. In the implementation process, each step of the above method or each module above can be completed by an integrated logic circuit of hardware in the processor element or instructions in the form of software.

[0104] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. Figure 6 As shown, the electronic device may include: a transceiver 121 , a processor 122 , and a memory 123 .

[0105] The processor 122 executes the computer execution instructions stored in the memory, so that the processor 122 executes the scheme in the above embodiment. The processor 122 can be a general-purpose processor, including a central processing unit CPU, a network processor (network processor, NP), etc.; it can also be a digital signal processor DSP, an application-specific integrated circuit ASIC, a field programmable gate array FPGA or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components.

[0106] The memory 123 is connected to the processor 122 via a system bus and completes communication between them. The memory 123 is used to store computer program instructions.

[0107] The transceiver 121 may be used to obtain tasks to be executed and configuration information of the tasks to be executed.

[0108] The system bus can be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus, etc. The system bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, only one thick line is used in the figure, but it does not mean that there is only one bus or one type of bus. The transceiver is used to realize the communication between the database access device and other computers (such as clients, read-write libraries, and read-only libraries). The memory may include random access memory (RAM) and may also include non-volatile memory.

[0109] The electronic device provided in the embodiment of the present application may be the terminal device of the above embodiment.

[0110] An embodiment of the present application also provides a computer-readable storage medium, in which computer instructions are stored. When the computer instructions are executed on a computer, the computer executes the technical solution of the small target detection method in the power scenario of the above embodiment.

[0111] An embodiment of the present application also provides a computer program product, which includes a computer program stored in a computer-readable storage medium. At least one processor can read the computer program from the computer-readable storage medium. When at least one processor executes the computer program, it can implement the technical solution of the small target detection method in the power scenario in the above embodiment.

[0112] In the process of implementation, the computer program product can be written in one or more programming languages ​​or a combination thereof to perform the computer program code of the present application, and the programming language includes an object-oriented programming language, such as Java, Smalltalk, C++, and also includes a conventional procedural programming language, such as "C" language or similar programming language. The program code can be executed entirely on the user's computer, partially on the user's computer, as an independent software package, partially on the user's computer and partially on the remote computer, or completely on the remote computer or server. In the case of a remote computer, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computer (for example, using an Internet service provider to connect through the Internet).

[0113] Note that the above are only preferred embodiments of the present application and the technical principles used. Those skilled in the art will understand that the present application is not limited to the specific embodiments herein, and that various obvious changes, readjustments and substitutions can be made by those skilled in the art without departing from the scope of protection of the present application. Therefore, although the present application is described in more detail through the above embodiments, the present application is not limited to the above embodiments, and may include more other equivalent embodiments without departing from the concept of the present application, and the scope of the present application is determined by the scope of the appended claims.

Claims

1. A small target detection method in an electric power scene, characterized in that: include: Input the image to be detected of the power scene into the first receptive field enhancement module group after passing through the first convolution layer to obtain a first enhanced feature map; Inputting the first enhanced feature map into a second receptive field enhancement module group to obtain a second enhanced feature map; Input the second enhanced feature map into the spatial pyramid pooling layer after passing through the second convolution layer to obtain a pooled feature map; Inputting the pooled feature map into a first multi-scale feature fusion module to obtain a first fused feature map; The first fusion feature map is input into a first detection head to obtain a first target detection result, wherein the first detection head is used to detect whether the image to be detected includes a small target of a preset first scale.

2. The method according to claim 1, characterized in that: Also includes: Inputting the pooled feature map into a third convolutional layer, and performing feature splicing on an output result of the third convolutional layer and the second enhanced feature map to obtain a first spliced ​​feature map; Inputting the first splicing feature into a second multi-scale feature fusion module to obtain a second fused feature map; The second fused feature map is input into a second detection head to obtain a second target detection result, wherein the second detection head is used to detect whether the image to be detected includes a small target of a preset second scale.

3. The method according to claim 2, characterized in that Also includes: Inputting the second fused feature map into a fourth convolutional layer, and performing feature splicing on an output result of the fourth convolutional layer and the first enhanced feature map to obtain a second spliced ​​feature map; Inputting the second splicing features into a third multi-scale feature fusion module to obtain a third fusion feature map; The third fusion feature map is input into a third detection head to obtain a third target detection result, wherein the third detection head is used to detect whether the image to be detected includes a small target of a preset third scale.

4. The method according to any one of claims 1 to 3, characterized in that The first receptive field enhancement module group and the second receptive field enhancement module group are each composed of two receptive field enhancement modules, and each receptive field enhancement module is used to expand the perception range of the image and capture features of different scales of the image.

5. The method according to claim 4, characterized in that The process of taking the input of any receptive field enhancement module as the current feature map and the receptive field enhancement module outputting the current enhanced feature map according to the current feature map includes: Inputting the current feature map into the fifth convolutional layer to obtain a first convolutional feature map, and inputting the first convolutional feature map into the sixth convolutional layer to obtain a second convolutional feature map; Performing feature splicing on the first convolution feature map and the second convolution feature map to obtain a third spliced ​​feature map; Passing the third concatenated feature map sequentially through a maximum pooling layer, an attention enhancement layer, a seventh convolutional layer, and an eighth convolutional layer to obtain a third convolutional feature map; Performing feature splicing on the first convolution feature map, the second convolution feature map, and the third convolution feature map to obtain a fourth spliced ​​feature map; The fourth spliced ​​feature map and the current feature map are feature spliced ​​to obtain the current enhanced feature map.

6. The method according to claim 5, characterized in that The feature concatenation method includes: adding feature elements, multiplying feature elements, or first adding feature elements and then multiplying them.

7. A small target detection device for power scene, characterized in that: include: A first enhancement module, used for inputting the image to be detected of the power scene into the first receptive field enhancement module group after passing through the first convolution layer, to obtain a first enhanced feature map; A second enhancement module, configured to input the first enhanced feature map into a second receptive field enhancement module group to obtain a second enhanced feature map; A pooling module, used for inputting the second enhanced feature map into a spatial pyramid pooling layer after passing through a second convolutional layer to obtain a pooled feature map; A first fusion module, used for inputting the pooled feature map into a first multi-scale feature fusion module to obtain a first fused feature map; The first detection module is used to input the first fusion feature map into a first detection head to obtain a first target detection result, wherein the first detection head is used to detect whether the image to be detected includes a small target of a preset first scale.

8. An electronic device, characterized in that: include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executable instructions; The processor executes the computer-executable instructions stored in the memory to implement the small target detection method for the power scenario according to any one of claims 1-6.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the small target detection method for an electric power scenario according to any one of claims 1 to 6.

10. A computer program product, characterized in that It comprises a computer program, which, when executed by a processor, implements the small target detection method for the power scenario described in any one of claims 1-6.

Images