Electric power image target detection method based on neural network architecture search and related device

Abstract

The invention belongs to the technical field of target detection, and discloses an electric power image target detection method based on neural network architecture search and a related device. The method comprises the following steps: acquiring a to-be-identified power image; inputting the power image into a pre-trained power image target detection model to obtain a detection result; and outputting the detection result. Wherein the pre-trained electric power image target detection model is an optimal neural network model obtained by training based on a joint evaluation index value. According to the method, effective training of the model is realized based on finite samples, and a random neural network search algorithm is utilized to update a model structure and parameters by taking a model discovery rate, a false alarm rate and a missing report rate as joint evaluation indexes; the technical problems that an existing target detection algorithm is low in discovery rate, high in false alarm rate and high in missing report rate are effectively solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of target detection, and in particular relates to a method for detecting power image targets based on a neural network architecture search and a related device. Background Art

[0002] With the increasing development of computer technology, artificial intelligence technology is increasingly widely used in power grid scenarios and plays an increasingly important role. However, the on-site environment of power operations is complex, the viewing angle of monitoring equipment is limited, and there are prominent problems such as mutual occlusion between equipment and personnel. The sample quality is low, and the artificial intelligence model cannot learn accurate classification features, resulting in low recognition accuracy. In addition, some visual samples are difficult to collect, and model training can only be carried out based on small-scale data sets. The artificial intelligence model cannot learn enough effective features and can only ensure the detection rate by increasing the number of iterations. It is easy to overfit, and all objects with similar features are identified as the same type, and interference objects cannot be correctly eliminated, resulting in increased false alarm and missed alarm rates of the model. Summary of the invention

[0003] The purpose of the present invention is to provide a power image target detection method and related devices based on neural network architecture search to solve the technical problems of low discovery rate, high false alarm rate and high missed alarm rate of existing target detection algorithms.

[0004] In order to achieve the above object, the present invention adopts the following technical solution:

[0005] The first invention provides a method for detecting power image targets based on a neural network architecture search, comprising:

[0006] Acquire an electric power image to be identified;

[0007] Inputting the power image into a pre-trained power image target detection model to obtain a detection result;

[0008] Outputting the detection result;

[0009] Among them, the pre-trained power image target detection model is an optimal neural network model obtained by training based on the joint evaluation index value.

[0010] The present invention is further improved in that: the power image target detection model includes CBL, Res1, Res2, Res8, Res8, Res4 connected in sequence, 5 CBLs connected in parallel, CBL and Conv.

[0011] The present invention is further improved in that: CBL is a basic convolution unit composed of one convolution layer Conv, one batch normalization layer BN and one activation function layer Leaky ReLU.

[0012] The present invention is further improved in that: ResX=Zero Padding+CBL+X*Res unit; wherein ZeroPadding represents zero padding; X represents X Res units, X=1, 2, 4 or 8; Res unit is a residual unit, and after the input passes through two CBLs, it is superimposed with the original input.

[0013] The present invention is further improved in that: in the step of obtaining the optimal neural network model based on the joint evaluation index training of the pre-trained power image target detection model, the training step of the power image target detection model includes:

[0014] Obtain a data set of power images and divide it into a training set, a validation set, and a test set;

[0015] The convolution kernel size in the convolution layer of each module of the power image target detection model is set to have two options: 1*1 and 3*3, and the number of convolution kernels is set to have three options: 32, 64, and 128. The convolution kernel size and the number of convolution kernels are combined in pairs to form six convolution layer structures, which constitute different CBL modules, and then form different ResX modules and the overall network structure to form the search space;

[0016] A network structure is randomly selected in the search space, the training set is input into the power image target detection model, the joint evaluation index value is calculated using the training set data, and the network parameters are updated based on the gradient descent method; then the validation set data is used to find the next network structure in the search space; the loop is repeated until the optimal network model is returned, the network model weight parameters and architecture parameters of the optimal network model are saved, and the trained power image target detection model is obtained.

[0017] The present invention is further improved in that: the joint evaluation index value is the inverse value of the recognition effect score;

[0018] The recognition effect score = discovery rate score + false detection rate score + missed detection rate score;

[0019] Among them, the discovery rate score = discovery rate × 60; discovery rate = M1 / M × 100%, M1 is the total number of correct boxes output by the power image object detection model, and M is the total number of standard boxes in the corresponding test image;

[0020] False detection ratio W = M2 / M, where M2 is the total number of error frames output by the power image target detection model; when W≤1, the false detection ratio score = 30 points; when 1<W≤2, the false detection ratio score = 25 points; when 2<W≤3, the false detection ratio score = 20 points; when 3<W≤4, the false detection ratio score = 10 points; when 4<W, the false detection ratio score = 0 points;

[0021] Missed detection rate score = (1-missed detection rate) × 10; Missed detection rate = M3 / M, where M3 is the total number of missed detection boxes of the power image object detection model.

[0022] In a second aspect, the present invention provides a power image target detection device based on neural network architecture search, comprising:

[0023] An acquisition module, used for acquiring the power image to be identified;

[0024] An object detection module is used to input the power image into a pre-trained power image object detection model to obtain a detection result;

[0025] An output module, used for outputting the detection result;

[0026] Among them, the pre-trained power image target detection model is an optimal neural network model obtained by training based on the joint evaluation index value.

[0027] The present invention is further improved in that: the power image target detection model includes CBL, Res1, Res2, Res8, Res8, Res4 connected in sequence, 5 CBLs connected in parallel, CBL and Conv.

[0028] The present invention is further improved in that: CBL is a basic convolution unit composed of one convolution layer Conv, one batch normalization layer BN and one activation function layer Leaky ReLU.

[0029] The present invention is further improved in that: ResX=Zero Padding+CBL+X*Res unit; wherein ZeroPadding represents zero padding; X represents X Res units, X=1, 2, 4 or 8; Res unit is a residual unit, and after the input passes through two CBLs, it is superimposed with the original input.

[0030] The present invention is further improved in that: in the step of obtaining the optimal neural network model based on the joint evaluation index training of the pre-trained power image target detection model, the training step of the power image target detection model includes:

[0031] Obtain a data set of power images and divide it into a training set, a validation set, and a test set;

[0032] The convolution kernel size in the convolution layer of each module of the power image target detection model is set to have two options: 1*1 and 3*3, and the number of convolution kernels is set to have three options: 32, 64, and 128. The convolution kernel size and the number of convolution kernels are combined in pairs to form six convolution layer structures, which constitute different CBL modules, and then form different ResX modules and the overall network structure to form the search space;

[0033] A network structure is randomly selected in the search space, the training set is input into the power image target detection model, the joint evaluation index value is calculated using the training set data, and the network parameters are updated based on the gradient descent method; then the validation set data is used to find the next network structure in the search space; the loop is repeated until the optimal network model is returned, the network model weight parameters and architecture parameters of the optimal network model are saved, and the trained power image target detection model is obtained.

[0034] The present invention is further improved in that: the joint evaluation index value is the inverse value of the recognition effect score;

[0035] The recognition effect score = discovery rate score + false detection rate score + missed detection rate score;

[0036] Among them, the discovery rate score = discovery rate × 60; discovery rate = M1 / M × 100%, M1 is the total number of correct boxes output by the power image object detection model, and M is the total number of standard boxes in the corresponding test image;

[0037] False detection ratio W = M2 / M, where M2 is the total number of error frames output by the power image target detection model; when W≤1, the false detection ratio score = 30 points; when 1<W≤2, the false detection ratio score = 25 points; when 2<W≤3, the false detection ratio score = 20 points; when 3<W≤4, the false detection ratio score = 10 points; when 4<W, the false detection ratio score = 0 points;

[0038] Missed detection rate score = (1-missed detection rate) × 10; Missed detection rate = M3 / M, where M3 is the total number of missed detection boxes of the power image object detection model.

[0039] In a third aspect, the present invention provides an electronic device, comprising:

[0040] A memory for storing executable instructions;

[0041] The processor is used to implement the power image target detection method based on neural network architecture search when executing the executable instructions stored in the memory.

[0042] In a fourth aspect, the present invention provides a computer-readable storage medium storing executable instructions for causing a processor to execute and implement the method for power image target detection based on neural network architecture search.

[0043] In a fifth aspect, the present invention provides a computer program product, comprising a computer program or instructions, which, when executed by a processor, implements the method for power image target detection based on neural network architecture search.

[0044] Compared with the prior art, the present invention has the following unexpected beneficial effects:

[0045] The present invention provides a method for detecting power image targets based on neural network architecture search, comprising: obtaining a power image to be identified; inputting the power image into a pre-trained power image target detection model to obtain a detection result; and outputting the detection result; wherein the pre-trained power image target detection model is an optimal neural network model obtained by training based on a joint evaluation index value. The present invention realizes effective training of the model based on finite samples, utilizes a random neural network search algorithm, and uses the model discovery rate, false alarm rate, and missed alarm rate as joint evaluation indicators to update the model structure and parameters.

[0046] Furthermore, the present invention applies a neural network search algorithm to the field of power image detection, which can provide a variety of alternative model structures for power image detection model training, realize automatic search of model parameters, and support continuous iterative optimization of the model.

[0047] Furthermore, the joint evaluation index based on model discovery rate, false alarm rate and missed alarm rate proposed in the present invention can achieve a more comprehensive evaluation of the model detection effect and avoid the situation of "high model discovery rate but high false alarm rate".

[0048] Furthermore, the model structure search method proposed in the present invention adopts a random search strategy, does not consume a large amount of computing resources, has a fast search speed, and can be used in scenarios where computing power resources are not highly configured.

[0049] Furthermore, the model structure search method proposed in the present invention can further expand the scope of the search space, providing a basis for achieving fully automatic optimization of the model structure. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] The accompanying drawings constituting a part of the present invention are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the accompanying drawings:

[0051] Figure 1 A schematic diagram of the structure of a power image target detection model provided by the present invention;

[0052] Figure 2 A structural diagram of a CBL and ResX module provided in an embodiment of the present invention;

[0053] Figure 3 A schematic diagram of a flow chart of a method for detecting power image targets based on a neural network architecture search provided by an embodiment of the present invention;

[0054] Figure 4 A schematic diagram of the structure of a power image target detection device based on neural network architecture search provided by an embodiment of the present invention;

[0055] Figure 5 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0056] The present invention will be described in detail below with reference to the accompanying drawings and in combination with embodiments. It should be noted that the embodiments and features in the embodiments of the present invention can be combined with each other without conflict.

[0057] The following detailed description is an exemplary description, which is intended to provide further detailed description of the present invention. Unless otherwise specified, all technical terms used in the present invention have the same meaning as those generally understood by those skilled in the art to which the present invention belongs. The terms used in the present invention are only for describing specific embodiments, and are not intended to limit the exemplary embodiments according to the present invention.

[0058] An embodiment of the present invention provides a method for detecting power image targets based on a neural network architecture search, comprising:

[0059] 1. Obtain power image samples; divide the power image samples into training set, validation set, and test set, where the ratio of training set to validation set is 1:1;

[0060] 2. Create CBL module and ResX module;

[0061] CBL: A basic convolution unit consisting of 1 convolution layer Conv, 1 batch normalization layer BN, and an activation function layer Leaky ReLU.

[0062] Res unit: A residual unit that adds the input after it passes through two CBLs. The purpose of the residual unit is to allow the network to extract deeper features while avoiding gradient vanishing or exploding.

[0063] ResX: ResX = Zero Padding + CBL + X * Res unit. Zero Padding means "zero padding", which refers to the technology of padding zeros at both ends of the data; X means X Res units.

[0064] In the convolutional layer of the above module, the convolution kernel size has two options: 1*1 and 3*3, and the number of convolution kernels has three options: 32, 64, and 128. The convolution kernel size and the number of convolution kernels are combined in pairs to form six convolution layer structures, which constitute different CBL modules, and then form different ResX modules and the overall network structure. The search space is thus constructed.

[0065] 3. Take CBL module and ResX module structure as the smallest unit, according to Figure 1The network structure shown is stacked to form a power image target detection model. In the convolution layer of the above module, there are two options for the convolution kernel size: 1*1 and 3*3, and three options for the number of convolution kernels: 32, 64, and 128. The selection of convolution kernel size and the number of convolution kernels are combined in pairs to form six convolution layer structures, which constitute different CBL modules, and then form different ResX modules and the overall network structure. The search space is thus formed. Specifically, the power image target detection model includes the following connected in sequence: CBL, Res1, Res2, Res8, Res8, Res4, 5 CBLs in parallel, CBL and Conv.

[0066] 4. Input the divided data set into the power image target detection model. First, randomly select a network structure from the search space, input the training set data and calculate the joint evaluation index according to the network output results, and update the network parameters based on the gradient descent method. The network structure after the parameter update is recorded as the candidate network; then use the validation set data to find the next network structure in the search space; repeat this step until the optimal network model is returned, save its network model weight parameters and architecture parameters, and obtain the trained power image target detection model.

[0067] In one embodiment, the validation set data is input into the candidate network structure, and the recognition effect score is calculated according to the network output result, which is recorded as the current optimal score. A network structure is randomly selected in the search space, and the validation set data is input into the network, and the recognition effect score is calculated according to the network output result. If the score of the network exceeds the current optimal score, it is used as a new network structure for training, otherwise the next network structure is randomly selected until the optimal score cannot be updated, and the current candidate network structure is the optimal network structure.

[0068] 5. The calculation process of the above joint evaluation index is as follows:

[0069] (1) Discovery rate = M1 / M × 100%. Among them, M1 is the total number of correct boxes output by the power image target detection model, and M is the total number of standard boxes in the corresponding test image. The basis for judging whether the algorithm output box is correct is: if the IOU value of the overlap rate between the output box area and the standard box area is not less than the specified threshold and the attribute description is correct, then the output box is correct. If the algorithm outputs multiple results for the same target position and the description is correct, the output box with the largest overlap rate is judged to be correct, and the other output boxes are judged to be incorrect.

[0070] (2) False detection ratio W = M2 / M. M2 is the total number of error boxes output by the power image target detection model, and M is the total number of standard boxes in the corresponding test image. The basis for judging whether the algorithm output box is wrong is as follows: the incorrect box output by the algorithm is the error box (i.e.: ① There is no standard box at the target position, but there is an output box; ② There is a standard box at the target position, but the IOU value of the overlap rate between the output box area and the standard box area is less than the specified threshold (IOU>0); ③ There is a standard box at the target position, but the output box category is wrong; ④ If the algorithm outputs multiple results for the same target position and the description is correct, the output box with the largest overlap rate is judged to be correct, and the other output boxes are judged to be wrong.

[0071] (3) Missed detection rate = M3 / M. M3 is the total number of missed detection frames of the power image target detection model, and M is the total number of standard frames in the corresponding test image. The basis for judging the missed detection frame of the algorithm is: there is a standard frame at the target position, but no output frame.

[0072] The discovery rate score and the false positive rate score are weighted, where the discovery rate score = discovery rate × 60, the false positive rate score is graded according to the following table, and the missed detection rate score = (1-missed detection rate) × 10. The calculation formula is as follows, with a full score of 100.

[0073] Identification effect score = discovery rate score (maximum score 60 points) + false detection rate score (maximum score 30 points) + missed detection rate score (maximum score 10 points);

[0074] False detection ratio (W) W≤1 1＜W≤2 2＜W≤3 3＜W≤4 4＜W False positive rate score 30 25 20 10 0

[0075] The inverse of the recognition effect score is taken as the joint evaluation index value to improve the accuracy of network structure search.

[0076] See also Figure 3 As shown, this embodiment provides a method for detecting power image targets based on neural network architecture search, including:

[0077] S1. Acquire a power image to be identified;

[0078] S2, inputting the power image into a pre-trained power image target detection model to obtain a detection result;

[0079] In a specific implementation, the pre-trained power image target detection model is an optimal neural network model obtained through training based on the joint evaluation index value.

[0080] In one specific embodiment, see Figure 1 As shown, the power image target detection model includes CBL, Res1, Res2, Res8, Res8, Res4 connected in sequence, 5 CBLs in parallel, CBL and Conv.

[0081] In a specific implementation, CBL is a basic convolution unit consisting of one convolution layer Conv, one batch normalization layer BN, and one activation function layer Leaky ReLU.

[0082] In a specific implementation, ResX=Zero Padding+CBL+X*Res unit; wherein Zero Padding represents zero padding; X represents X Res units, X=1, 2, 4 or 8; Res unit is a residual unit, and after the input passes through two CBLs, it is superimposed with the original input.

[0083] In a specific implementation, the training steps of the power image target detection model include:

[0084] Obtain a data set of power images and divide it into a training set, a validation set, and a test set;

[0085] The convolution kernel size in the convolution layer of each module of the power image target detection model is set to have two options: 1*1 and 3*3, and the number of convolution kernels is set to have three options: 32, 64, and 128. The convolution kernel size and the number of convolution kernels are combined in pairs to form six convolution layer structures, which constitute different CBL modules, and then form different ResX modules and the overall network structure to form the search space;

[0086] A network structure is randomly selected in the search space, the training set is input into the power image target detection model, the joint evaluation index value is calculated using the training set data, and the network parameters are updated based on the gradient descent method; then the validation set data is used to find the next network structure in the search space; the loop is repeated until the optimal network model is returned, the network model weight parameters and architecture parameters of the optimal network model are saved, and the trained power image target detection model is obtained.

[0087] The joint evaluation index value is the inverse value of the recognition effect score;

[0088] The recognition effect score = discovery rate score + false detection rate score + missed detection rate score;

[0089] Among them, the discovery rate score = discovery rate × 60; discovery rate = M1 / M × 100%, M1 is the total number of correct boxes output by the power image object detection model, and M is the total number of standard boxes in the corresponding test image;

[0090] False detection ratio W = M2 / M, where M2 is the total number of error frames output by the power image target detection model; when W≤1, the false detection ratio score = 30 points; when 1<W≤2, the false detection ratio score = 25 points; when 2<W≤3, the false detection ratio score = 20 points; when 3<W≤4, the false detection ratio score = 10 points; when 4<W, the false detection ratio score = 0 points;

[0091] Missed detection rate score = (1-missed detection rate) × 10; Missed detection rate = M3 / M, where M3 is the total number of missed detection boxes of the power image object detection model.

[0092] S3. Output the detection result.

[0093] See also Figure 4 As shown, an embodiment of the present invention provides a power image target detection device based on neural network architecture search, comprising:

[0094] An acquisition module, used for acquiring the power image to be identified;

[0095] An object detection module is used to input the power image into a pre-trained power image object detection model to obtain a detection result;

[0096] An output module, used for outputting the detection result;

[0097] Among them, the pre-trained power image target detection model is an optimal neural network model obtained by training based on the joint evaluation index value.

[0098] In a specific implementation, the power image target detection model includes CBL, Res1, Res2, Res8, Res8, Res4 connected in sequence, 5 CBLs connected in parallel, CBL and Conv.

[0099] In a specific implementation, CBL is a basic convolution unit consisting of one convolution layer Conv, one batch normalization layer BN, and one activation function layer Leaky ReLU.

[0100] In a specific implementation, ResX=Zero Padding+CBL+X*Res unit; wherein Zero Padding represents zero padding; X represents X Res units, X=1, 2, 4 or 8; Res unit is a residual unit, and after the input passes through two CBLs, it is superimposed with the original input.

[0101] In a specific implementation, in the step of obtaining the optimal neural network model based on the joint evaluation index training of the pre-trained power image target detection model, the training step of the power image target detection model includes:

[0102] Obtain a data set of power images and divide it into a training set, a validation set, and a test set;

[0103] The convolution kernel size in the convolution layer of each module of the power image target detection model is set to have two options: 1*1 and 3*3, and the number of convolution kernels is set to have three options: 32, 64, and 128. The convolution kernel size and the number of convolution kernels are combined in pairs to form six convolution layer structures, which constitute different CBL modules, and then form different ResX modules and the overall network structure to form the search space;

[0104] A network structure is randomly selected in the search space, the training set is input into the power image target detection model, the joint evaluation index value is calculated using the training set data, and the network parameters are updated based on the gradient descent method; then the validation set data is used to find the next network structure in the search space; the loop is repeated until the optimal network model is returned, the network model weight parameters and architecture parameters of the optimal network model are saved, and the trained power image target detection model is obtained.

[0105] In a specific implementation, the joint evaluation index value is the inverse value of the recognition effect score;

[0106] The recognition effect score = discovery rate score + false detection rate score + missed detection rate score;

[0107] Among them, the discovery rate score = discovery rate × 60; discovery rate = M1 / M × 100%, M1 is the total number of correct boxes output by the power image object detection model, and M is the total number of standard boxes in the corresponding test image;

[0108] False detection ratio W = M2 / M, where M2 is the total number of error frames output by the power image target detection model; when W≤1, the false detection ratio score = 30 points; when 1<W≤2, the false detection ratio score = 25 points; when 2<W≤3, the false detection ratio score = 20 points; when 3<W≤4, the false detection ratio score = 10 points; when 4<W, the false detection ratio score = 0 points;

[0109] Missed detection rate score = (1-missed detection rate) × 10; Missed detection rate = M3 / M, where M3 is the total number of missed detection boxes of the power image object detection model.

[0110] See also Figure 4 As shown, an embodiment of the present invention provides an electronic device 100 for implementing a method for detecting power image targets based on a neural network architecture search; the electronic device 100 includes a memory 101, at least one processor 102, a computer program 103 stored in the memory 101 and executable on the at least one processor 102, and at least one communication bus 104.

[0111] The memory 101 can be used to store the computer program 103, and the processor 102 implements the steps of the power image target detection method based on neural network architecture search by running or executing the computer program stored in the memory 101 and calling the data stored in the memory 101. The memory 101 can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, an application required for at least one function (such as a sound playback function, an image playback function, etc.), etc.; the data storage area can store data (such as audio data) created according to the use of the electronic device 100, etc. In addition, the memory 101 can include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other non-volatile solid-state storage devices.

[0112] The at least one processor 102 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The processor 102 may be a microprocessor or any conventional processor, etc. The processor 102 is the control center of the electronic device 100, and uses various interfaces and lines to connect various parts of the entire electronic device 100.

[0113] The memory 101 in the electronic device 100 stores a plurality of instructions to implement a method for detecting power image targets based on a neural network architecture search, and the processor 102 can execute the plurality of instructions to implement:

[0114] Acquire an electric power image to be identified;

[0115] Inputting the power image into a pre-trained power image target detection model to obtain a detection result;

[0116] Outputting the detection result;

[0117] Among them, the pre-trained power image target detection model is an optimal neural network model obtained by training based on the joint evaluation index value.

[0118] If the module / unit integrated in the electronic device 100 is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present invention implements all or part of the process in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned method embodiments when executed by the processor. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may include: any entity or device that can carry the computer program code, recording medium, U disk, mobile hard disk, disk, optical disk, computer memory and read-only memory (ROM, Read-Only Memory).

[0119] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0120] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0121] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0122] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0123] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for power image target detection based on neural network architecture search, characterized in that: include: Acquire an electric power image to be identified; Inputting the power image into a pre-trained power image target detection model to obtain a detection result; Outputting the detection result; Among them, the pre-trained power image target detection model is an optimal neural network model obtained by training based on the joint evaluation index value.

2. The method for detecting power image targets based on neural network architecture search according to claim 1, characterized in that: The power image target detection model includes CBL, Res1, Res2, Res8, Res8, Res4 connected in sequence, 5 CBLs connected in parallel, CBL and Conv.

3. The method for detecting power image targets based on neural network architecture search according to claim 2, characterized in that: CBL is a basic convolution unit consisting of one convolution layer Conv, one batch normalization layer BN, and an activation function layer Leaky ReLU.

4. The method for detecting power image targets based on neural network architecture search according to claim 2, characterized in that: ResX=Zero Padding+CBL+X*Res unit; where Zero Padding represents zero padding; X represents X Res units, X=1, 2, 4 or 8; Res unit is a residual unit, and after the input passes through two CBLs, it is superimposed with the original input.

5. The method for detecting power image targets based on neural network architecture search according to claim 4, characterized in that: In the step of obtaining the optimal neural network model based on the joint evaluation index training of the pre-trained power image target detection model, the training step of the power image target detection model includes: Obtain a data set of power images and divide it into a training set, a validation set, and a test set; The convolution kernel size in the convolution layer of each module of the power image target detection model is set to have two options: 1*1 and 3*3, and the number of convolution kernels is set to have three options: 32, 64, and 128. The convolution kernel size and the number of convolution kernels are combined in pairs to form six convolution layer structures, which constitute different CBL modules, and then form different ResX modules and the overall network structure to form the search space; A network structure is randomly selected in the search space, the training set is input into the power image target detection model, the joint evaluation index value is calculated using the training set data, and the network parameters are updated based on the gradient descent method; then the validation set data is used to find the next network structure in the search space; the loop is repeated until the optimal network model is returned, the network model weight parameters and architecture parameters of the optimal network model are saved, and the trained power image target detection model is obtained.

6. The method for detecting power image targets based on neural network architecture search according to claim 1, characterized in that: The joint evaluation index value is the inverse value of the recognition effect score; The recognition effect score = discovery rate score + false detection rate score + missed detection rate score; Among them, the discovery rate score = discovery rate × 60; discovery rate = M1 / M × 100%, M1 is the total number of correct boxes output by the power image object detection model, and M is the total number of standard boxes in the corresponding test image; False detection ratio W = M2 / M, where M2 is the total number of error frames output by the power image target detection model; when W≤1, the false detection ratio score = 30 points; when 1<W≤2, the false detection ratio score = 25 points; when 2<W≤3, the false detection ratio score = 20 points; when 3<W≤4, the false detection ratio score = 10 points; when 4<W, the false detection ratio score = 0 points; Missed detection rate score = (1-missed detection rate) × 10; Missed detection rate = M3 / M, where M3 is the total number of missed detection boxes of the power image object detection model.

7. A power image target detection device based on neural network architecture search, characterized in that: include: An acquisition module, used for acquiring the power image to be identified; An object detection module is used to input the power image into a pre-trained power image object detection model to obtain a detection result; An output module, used for outputting the detection result; Among them, the pre-trained power image target detection model is an optimal neural network model obtained by training based on the joint evaluation index value.

8. The power image target detection device based on neural network architecture search according to claim 7, characterized in that: The power image target detection model includes CBL, Res1, Res2, Res8, Res8, Res4 connected in sequence, 5 CBLs connected in parallel, CBL and Conv.

9. The power image target detection device based on neural network architecture search according to claim 8, characterized in that: CBL is a basic convolution unit consisting of one convolution layer Conv, one batch normalization layer BN, and an activation function layer Leaky ReLU.

10. The power image target detection device based on neural network architecture search according to claim 8, characterized in that: ResX=Zero Padding+CBL+X*Res unit; where Zero Padding represents zero padding; X represents X Res units, X=1, 2, 4 or 8; Res unit is a residual unit, and after the input passes through two CBLs, it is superimposed with the original input.

11. The power image target detection device based on neural network architecture search according to claim 10, characterized in that: In the step of obtaining the optimal neural network model based on the joint evaluation index training of the pre-trained power image target detection model, the training step of the power image target detection model includes: Obtain a data set of power images and divide it into a training set, a validation set, and a test set; The convolution kernel size in the convolution layer of each module of the power image target detection model is set to have two options: 1*1 and 3*3, and the number of convolution kernels is set to have three options: 32, 64, and 128. The convolution kernel size and the number of convolution kernels are combined in pairs to form six convolution layer structures, which constitute different CBL modules, and then form different ResX modules and the overall network structure to form the search space; A network structure is randomly selected in the search space, the training set is input into the power image target detection model, the joint evaluation index value is calculated using the training set data, and the network parameters are updated based on the gradient descent method; then the validation set data is used to find the next network structure in the search space; the loop is repeated until the optimal network model is returned, the network model weight parameters and architecture parameters of the optimal network model are saved, and the trained power image target detection model is obtained.

12. The power image target detection device based on neural network architecture search according to claim 7, characterized in that: The joint evaluation index value is the inverse value of the recognition effect score; The recognition effect score = discovery rate score + false detection rate score + missed detection rate score; Among them, the discovery rate score = discovery rate × 60; discovery rate = M1 / M × 100%, M1 is the total number of correct boxes output by the power image object detection model, and M is the total number of standard boxes in the corresponding test image; False detection ratio W = M2 / M, where M2 is the total number of error frames output by the power image target detection model; when W≤1, the false detection ratio score = 30 points; when 1<W≤2, the false detection ratio score = 25 points; when 2<W≤3, the false detection ratio score = 20 points; when 3<W≤4, the false detection ratio score = 10 points; when 4<W, the false detection ratio score = 0 points; Missed detection rate score = (1-missed detection rate) × 10; Missed detection rate = M3 / M, where M3 is the total number of missed detection boxes of the power image object detection model.

13. An electronic device, characterized in that: include: A memory for storing executable instructions; A processor, configured to implement the method according to any one of claims 1 to 6 when executing the executable instructions stored in the memory.

14. A computer-readable storage medium, characterized in that: Executable instructions are stored, which are used to cause a processor to execute and implement the method described in any one of claims 1 to 6.

15. A computer program product comprising a computer program or instructions, characterized in that When the computer program or instruction is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images