Method and device for intelligently identifying equipment

Abstract

The invention relates to the technical field of equipment identification, and particularly discloses a method and a device for intelligently identifying equipment, and the method comprises the steps: collecting a target image in a machine room inspection scene; based on a predetermined positioning identification model, analyzing the target image to obtain an actual position of each device in the machine room inspection scene and a device type of the device; based on a predetermined classification identification model, analyzing the positioning identification result and the target image of each device in the machine room inspection scene to obtain the component category of the sub-component in each device in the machine room inspection scene, according to the method, the positioning and identification accuracy of equipment and sub-components (such as ports and clamping plates) in the equipment with similar appearance characteristics or port appearance characteristics is improved, accurate and efficient distinguishing between the equipment and between the sub-components in the equipment is realized, and stable positioning and identification can be realized under the inspection condition with low calculation power.

Description

Technical Field

[0001] The present invention relates to the technical field of device identification, and in particular to a method and device for intelligently identifying a device. Background Art

[0002] The existing computer room inspection system mainly relies on manual visual inspection to determine the status of the equipment in the computer room, but it is inefficient and inaccurate. Therefore, a machine vision-based inspection method came into being. Specifically, the image of the equipment in the computer room is collected by the camera, and the collected image is analyzed using the traditional mainstream machine vision model (such as Faster R-CNN, template matching, Hough transform), and then the equipment in the computer room is located and identified based on the image analysis results.

[0003] However, practice has found that the existing equipment positioning and identification methods cannot accurately locate and identify equipment with similar appearance features or port appearance features, and still require manual intervention, resulting in low inspection efficiency and accuracy. Therefore, it is very necessary to propose a technical solution to improve the accuracy of positioning and identification of equipment with similar appearance features or port appearance features. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a method and device for intelligently identifying devices, which can improve the accuracy of positioning and identifying devices with similar appearance features or port appearance features.

[0005] In order to solve the above technical problems, the first aspect of the present invention discloses a method for intelligently identifying a device, which is applied in a computer room inspection scenario, and comprises:

[0006] Collecting a target image in a computer room inspection scene, where there are multiple devices in the computer room inspection scene, and each of the devices includes multiple sub-components;

[0007] Based on a predetermined positioning recognition model, the target image is analyzed to obtain a positioning recognition result of each of the devices in the computer room inspection scene, wherein the positioning recognition result of each of the devices in the computer room inspection scene includes the actual position of the device in the computer room inspection scene and the device category of the device, wherein the positioning recognition model is obtained by training based on the YOLOv7 basic positioning model;

[0008] Based on a predetermined classification and recognition model, the positioning and recognition results of each device in the computer room inspection scene and the target image are analyzed to obtain a category recognition result of the computer room inspection scene, wherein the classification and recognition model is obtained by training based on the SVM basic classification and recognition model;

[0009] The category identification result of the computer room inspection scene includes the component category and the position of the subcomponent of each subcomponent in the equipment in the computer room inspection scene.

[0010] As an optional implementation, in the first aspect of the present invention, the target image is analyzed based on a predetermined positioning and recognition model to obtain a positioning and recognition result of each device in the computer room inspection scene, including:

[0011] Extracting spatial hierarchical information of the target image based on a backbone network of a predetermined positioning recognition model, and analyzing the spatial hierarchical information of the target image to obtain a plurality of feature maps of the target image, wherein the size of each feature map in the target image is different;

[0012] Based on the neck network of the positioning recognition model, a fusion operation is performed on multiple feature maps of the target image to obtain a fusion result corresponding to the target image;

[0013] The fusion result corresponding to the target image and multiple feature maps of the target image are analyzed according to the head network of the positioning and recognition model to obtain the positioning and recognition results of each of the devices in the computer room inspection scene.

[0014] As an optional implementation, in the first aspect of the present invention, the head network according to the positioning and recognition model analyzes the fusion result of the target image and the multiple feature maps of the target image to obtain the positioning and recognition result of each of the devices in the machine room inspection scene, including:

[0015] Performing prediction frame analysis on the fusion result of the target image according to the head network of the positioning recognition model to obtain multiple device prediction frames corresponding to the target image and the confidence of each of the device prediction frames;

[0016] For any of the device prediction frames, comparing the confidence of the device prediction frame with a preset confidence to obtain a confidence comparison result;

[0017] According to the confidence comparison result corresponding to each of the device prediction frames, all target device prediction frames whose confidence is greater than or equal to the preset confidence are screened out from all the device prediction frames;

[0018] According to the head network of the positioning recognition model, using multiple feature maps of the target image, performing category analysis on all the target device prediction boxes to obtain a category result of each of the target device prediction boxes, wherein the category result of each of the target device prediction boxes includes a probability of at least one category;

[0019] For any of the target device prediction boxes, the category with the largest probability is selected from the probabilities of all categories contained in the category results of the target device prediction box as the device category to which the target device prediction box belongs, and the actual position corresponding to the position of the target device prediction box in the target image in the computer room inspection scene is determined to be the actual position corresponding to the device corresponding to the device category to which the target device prediction box belongs in the computer room inspection scene.

[0020] As an optional embodiment, in the first aspect of the present invention, the method further comprises:

[0021] Acquire first device image training data in a sample computer room inspection scene, wherein the first device image training data includes a position and a device category of a device in each of a plurality of first sample device images in the sample computer room inspection scene;

[0022] Performing a training operation on the YOLOv7 basic positioning model using the first device image training data;

[0023] In the process of training the YOLOv7 basic positioning model, based on the predetermined loss network, calculating the loss parameter of the YOLOv7 basic positioning model in the training process;

[0024] According to the loss parameter corresponding to the YOLOv7 basic positioning model, an optimization operation is performed on the YOLOv7 basic positioning model in the training process to obtain the optimized YOLOv7 basic positioning model;

[0025] The optimized YOLOv7 basic positioning model is determined as the positioning recognition model.

[0026] As an optional embodiment, in the first aspect of the present invention, the method further comprises:

[0027] Acquire second device image training data in a sample computer room inspection scene, wherein the second device image training data includes a position and a category corresponding to each second device sample image in a plurality of second sample device images, each second device sample image includes at least one second sample device, each second sample device includes at least three sample sub-components, the position corresponding to each second device sample image includes the device positions of all the second sample devices and the positions of the sub-components included in each second sample device, and the category corresponding to each second device sample image includes the component category of the sub-components included in each second sample device;

[0028] Based on the second device image training data, a training operation is performed on the predetermined SVM basic classification recognition model until a model optimization parameter that satisfies the predetermined model optimization condition is obtained;

[0029] The model optimization parameters that meet the model optimization conditions are updated to the SVM basic classification recognition model of this training, and the updated SVM basic classification recognition model is determined as the classification recognition model.

[0030] As an optional implementation, in the first aspect of the present invention, the performing of a training operation on a predetermined SVM basic classification recognition model based on the second device image training data until obtaining model optimization parameters that meet predetermined model optimization conditions includes:

[0031] According to the sample computer room inspection scenario, a preset number of target model parameters are screened from the model parameters of the predetermined gray wolf optimization model, wherein, for any second sample device image, the number of all the sample subcomponents in the second sample device image is greater than or equal to the preset number, and the target model parameters corresponding to the second sample device image are not repeated;

[0032] For any second sample device image in the second device image training data, based on the position of the sample subcomponent contained in the second sample device image and the category of the sample subcomponent, all the target model parameters of the gray wolf optimization model are analyzed to obtain data of candidate points matching the second sample device image, wherein the data of the candidate points corresponding to the second sample device image include the position of the candidate point and the distance between the candidate point and each of the sample subcomponents in the second sample device image;

[0033] Based on the target data corresponding to the second device image training data, the predetermined SVM basic classification recognition model is trained until a model optimization parameter that satisfies the predetermined model optimization condition is obtained;

[0034] The target data corresponding to the second device image training data includes the position of the sample sub-component contained in each of the second sample device images, the component category of the sample sub-component, the position of the candidate point, the distance between the candidate point and each sample sub-component, and the device position.

[0035] As an optional implementation, in the first aspect of the present invention, the model optimization condition includes that the number of training iterations is greater than or equal to a preset number of iterations and / or the model optimization parameters obtained after each training are within a predetermined parameter range;

[0036] The parameter range is determined based on one or more of the density of equipment in the sample computer room inspection scene, the type of equipment, the density of sample sub-components in the equipment, the size of the sample sub-components, and the shape of the sample sub-components.

[0037] A second aspect of the present invention discloses a device for intelligently identifying a device, the device being applied in a computer room inspection scenario, and the device comprising:

[0038] A collection module, used for collecting target images in a computer room inspection scene, wherein there are multiple devices in the computer room inspection scene, and each of the devices includes multiple sub-components;

[0039] An analysis module is used to analyze the target image based on a predetermined positioning recognition model to obtain a positioning recognition result of each of the devices in the computer room inspection scene, wherein the positioning recognition result of each of the devices in the computer room inspection scene includes the actual position of the device in the computer room inspection scene and the device category of the device, wherein the positioning recognition model is obtained by training based on the YOLOv7 basic positioning model;

[0040] The analysis module is further used to analyze the positioning recognition results of each of the devices in the computer room inspection scene and the target image based on a predetermined classification recognition model to obtain a category recognition result of the computer room inspection scene, wherein the classification recognition model is obtained by training based on the SVM basic classification recognition model;

[0041] The category identification result of the computer room inspection scene includes the component category and the position of the subcomponent of each subcomponent in the equipment in the computer room inspection scene.

[0042] As an optional implementation, in the second aspect of the present invention, the analysis module analyzes the target image based on a predetermined positioning and recognition model, and a specific method for obtaining a positioning and recognition result of each of the devices in the computer room inspection scene includes:

[0043] Extracting spatial hierarchical information of the target image based on a backbone network of a predetermined positioning recognition model, and analyzing the spatial hierarchical information of the target image to obtain a plurality of feature maps of the target image, wherein the size of each feature map in the target image is different;

[0044] Based on the neck network of the positioning recognition model, a fusion operation is performed on multiple feature maps of the target image to obtain a fusion result corresponding to the target image;

[0045] The fusion result corresponding to the target image and multiple feature maps of the target image are analyzed according to the head network of the positioning and recognition model to obtain the positioning and recognition results of each of the devices in the computer room inspection scene.

[0046] As an optional implementation, in the second aspect of the present invention, the analysis module analyzes the fusion result of the target image and the multiple feature maps of the target image according to the head network of the positioning and recognition model to obtain the specific method of the positioning and recognition result of each of the devices in the computer room inspection scene, including:

[0047] Performing prediction frame analysis on the fusion result of the target image according to the head network of the positioning recognition model to obtain multiple device prediction frames corresponding to the target image and the confidence of each of the device prediction frames;

[0048] For any of the device prediction frames, comparing the confidence of the device prediction frame with a preset confidence to obtain a confidence comparison result;

[0049] According to the confidence comparison result corresponding to each of the device prediction frames, all target device prediction frames whose confidence is greater than or equal to the preset confidence are screened out from all the device prediction frames;

[0050] According to the head network of the positioning recognition model, using multiple feature maps of the target image, performing category analysis on all the target device prediction boxes to obtain a category result of each of the target device prediction boxes, wherein the category result of each of the target device prediction boxes includes a probability of at least one category;

[0051] For any of the target device prediction boxes, the category with the largest probability is selected from the probabilities of all categories contained in the category results of the target device prediction box as the device category to which the target device prediction box belongs, and the actual position corresponding to the position of the target device prediction box in the target image in the computer room inspection scene is determined to be the actual position corresponding to the device corresponding to the device category to which the target device prediction box belongs in the computer room inspection scene.

[0052] As an optional implementation, in the second aspect of the present invention, the device further includes:

[0053] A first acquisition module is used to acquire first device image training data in a sample computer room inspection scene, wherein the first device image training data includes a position and a device category of a device in each of a plurality of first sample device images in the sample computer room inspection scene;

[0054] A first training module, configured to perform a training operation on the YOLOv7 basic positioning model using the first device image training data;

[0055] A calculation module, used for calculating the loss parameters of the YOLOv7 basic positioning model in the training process based on a predetermined loss network during the training of the YOLOv7 basic positioning model;

[0056] An optimization module, used to perform an optimization operation on the YOLOv7 basic positioning model in the training process according to the loss parameter corresponding to the YOLOv7 basic positioning model, so as to obtain the optimized YOLOv7 basic positioning model;

[0057] The first determination module is used to determine the optimized YOLOv7 basic positioning model as the positioning recognition model.

[0058] As an optional implementation, in the second aspect of the present invention, the device further includes:

[0059] A second acquisition module is used to acquire second device image training data in a sample computer room inspection scene, wherein the second device image training data includes a position and a category corresponding to each second device sample image in a plurality of second sample device images, each second device sample image includes at least one second sample device, each second sample device includes at least three sample sub-components, the position corresponding to each second device sample image includes the device positions of all the second sample devices and the positions of the sub-components included in each second sample device, and the category corresponding to each second device sample image includes the component category of the sub-components included in each second sample device;

[0060] A second training module, used to perform a training operation on a predetermined SVM basic classification recognition model based on the second device image training data, until a model optimization parameter satisfying a predetermined model optimization condition is obtained;

[0061] The second determination module is used to update the model optimization parameters that meet the model optimization conditions to the SVM basic classification recognition model of this training, and determine the updated SVM basic classification recognition model as the classification recognition model.

[0062] As an optional implementation, in the second aspect of the present invention, the second training module performs a training operation on a predetermined SVM basic classification recognition model based on the second device image training data until a model optimization parameter that satisfies a predetermined model optimization condition is obtained. The specific method includes:

[0063] According to the sample computer room inspection scenario, a preset number of target model parameters are screened from the model parameters of the predetermined gray wolf optimization model, wherein, for any second sample device image, the number of all the sample subcomponents in the second sample device image is greater than or equal to the preset number, and the target model parameters corresponding to the second sample device image are not repeated;

[0064] For any second sample device image in the second device image training data, based on the position of the sample subcomponent contained in the second sample device image and the category of the sample subcomponent, all the target model parameters of the gray wolf optimization model are analyzed to obtain data of candidate points matching the second sample device image, wherein the data of the candidate points corresponding to the second sample device image include the position of the candidate point and the distance between the candidate point and each of the sample subcomponents in the second sample device image;

[0065] Based on the target data corresponding to the second device image training data, the predetermined SVM basic classification recognition model is trained until a model optimization parameter that satisfies the predetermined model optimization condition is obtained;

[0066] The target data corresponding to the second device image training data includes the position of the sample sub-component contained in each of the second sample device images, the component category of the sample sub-component, the position of the candidate point, the distance between the candidate point and each sample sub-component, and the device position.

[0067] As an optional implementation, in the second aspect of the present invention, the model optimization condition includes that the number of training iterations is greater than or equal to a preset number of iterations and / or the model optimization parameters obtained after each training are within a predetermined parameter range;

[0068] The parameter range is determined based on one or more of the density of equipment in the sample computer room inspection scene, the type of equipment, the density of sample sub-components in the equipment, the size of the sample sub-components, and the shape of the sample sub-components.

[0069] The third aspect of the present invention discloses another device for intelligently identifying a device, the device comprising:

[0070] A memory storing executable program code;

[0071] a processor coupled to the memory;

[0072] The processor calls the executable program code stored in the memory to execute the method for intelligently identifying a device disclosed in the first aspect of the present invention.

[0073] The fourth aspect of the present invention discloses a computer-readable storage medium, which stores computer instructions. When the computer instructions are called, they are used to execute the method for intelligently identifying a device disclosed in the first aspect of the present invention.

[0074] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0075] In an embodiment of the present invention, a target image in a computer room inspection scene is collected, and there are multiple devices in the computer room inspection scene, and each device includes multiple sub-components; based on a predetermined positioning and recognition model, the target image is analyzed to obtain a positioning and recognition result of each device in the computer room inspection scene, and the positioning and recognition result of each device in the computer room inspection scene includes the actual position of the device in the computer room inspection scene and the device category of the device, wherein the positioning and recognition model is obtained by training based on the YOLOv7 basic positioning model; based on a predetermined classification and recognition model, the positioning and recognition result of each device in the computer room inspection scene and the target image are analyzed to obtain a category recognition result of the computer room inspection scene, wherein the classification and recognition model is obtained by training based on the SVM basic classification and recognition model; wherein the category recognition result of the computer room inspection scene includes the component category of the sub-component in each device in the computer room inspection scene and the position of each sub-component, thereby improving the positioning and recognition accuracy of devices with similar appearance features or port appearance features and sub-components in the devices (such as ports, card boards, etc.), realizing accurate and efficient distinction between devices and between sub-components in devices, and can stably locate and recognize under low computing power inspection conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0077] Figure 1 It is a flowchart of a method for intelligently identifying a device disclosed in an embodiment of the present invention;

[0078] Figure 2 It is a schematic diagram of intelligent identification of a device disclosed in the prior art;

[0079] Figure 3 It is a schematic diagram of an embodiment of the present invention for applying the present invention to intelligently identify a device;

[0080] Figure 4 It is a structural schematic diagram of a device for intelligently identifying a device disclosed in an embodiment of the present invention;

[0081] Figure 5 It is a structural schematic diagram of another device for intelligently identifying a device disclosed in an embodiment of the present invention;

[0082] Figure 6 It is a structural schematic diagram of another apparatus for intelligently identifying a device disclosed in an embodiment of the present invention. DETAILED DESCRIPTION

[0083] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0084] The terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish different objects rather than to describe a specific order. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, device, product or end including a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units that are not listed, or may optionally include other steps or units inherent to these processes, methods, products or ends.

[0085] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present invention. The appearance of the phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0086] The present invention discloses a method and device for intelligently identifying equipment. The method collects target images in a computer room inspection scene. There are multiple devices in the computer room inspection scene, and each device includes multiple sub-components. Based on a predetermined positioning recognition model, the target image is analyzed to obtain a positioning recognition result of each device in the computer room inspection scene. The positioning recognition result of each device in the computer room inspection scene includes the actual position of the device in the computer room inspection scene and the device category of the device. The positioning recognition model is obtained by training based on the YOLOv7 basic positioning model. Based on a predetermined classification recognition model, the computer room inspection scene is The positioning and recognition results of each device in the inspection scene and the target image are analyzed to obtain the category recognition results of the computer room inspection scene, wherein the classification recognition model is obtained by training based on the SVM basic classification recognition model; wherein, the category recognition results of the computer room inspection scene include the component category and position of each sub-component in each device in the computer room inspection scene, which improves the positioning and recognition accuracy of devices with similar appearance features or port appearance features and sub-components in the device (such as ports, card boards, etc.), realizes accurate and efficient distinction between devices and sub-components in the device, and can stably locate and recognize under low computing power inspection conditions. The following are detailed descriptions.

[0087] Embodiment 1

[0088] See also Figure 1 , Figure 1 1 is a flow chart of a method for intelligently identifying a device disclosed in an embodiment of the present invention. Figure 1 The method for intelligently identifying devices described above can be applied to any scenario where it is necessary to identify the location and category of devices and their sub-components, such as a computer room inspection scenario. Figure 1 As shown, the method may include the following operations:

[0089] 101. Collect a target image in a computer room inspection scene. There are multiple devices in the computer room inspection scene, and each device includes multiple sub-components.

[0090] In the embodiment of the present invention, optionally, the subcomponents included in each device include but are not limited to one or more of ports, boards, card slots, wiring, and labels, etc. Optionally, the target image in the inspection scene of the computer room can be collected by a fixed camera, or by a movable camera.

[0091] 102. Based on a predetermined positioning and recognition model, the target image is analyzed to obtain the positioning and recognition results of each device in the computer room inspection scene. The positioning and recognition results of each device in the computer room inspection scene include the actual position of the device in the computer room inspection scene and the device category of the device. The positioning and recognition model is trained based on the YOLOv7 basic positioning model (such as YOLOv7-tiny).

[0092] 103. Based on a predetermined classification and recognition model, the positioning and recognition results and target images of each device in the computer room inspection scene are analyzed to obtain a category recognition result of the computer room inspection scene, wherein the classification and recognition model is trained based on the SVM basic classification and recognition model; wherein the category recognition result of the computer room inspection scene includes the component category of the sub-components in each device in the computer room inspection scene and the position of each sub-component.

[0093] In the embodiment of the present invention, further, each subcomponent may include one or more of color, shape and size.

[0094] It can be seen that the present invention is implemented to collect target images in the computer room inspection scene. There are multiple devices in the computer room inspection scene, and each device includes multiple sub-components; based on a predetermined positioning and recognition model, the target image is analyzed to obtain the positioning and recognition results of each device in the computer room inspection scene, and the positioning and recognition results of each device in the computer room inspection scene include the actual position of the device in the computer room inspection scene and the device category of the device, wherein the positioning and recognition model is obtained by training based on the YOLOv7 basic positioning model; based on a predetermined classification and recognition model, the positioning and recognition results and target images of each device in the computer room inspection scene are analyzed to obtain the category recognition results of the computer room inspection scene, wherein the classification and recognition model is obtained by training based on the SVM basic classification and recognition model; wherein the category recognition results of the computer room inspection scene include the component category of the sub-components in each device in the computer room inspection scene and the position of each sub-component, which improves the positioning and recognition accuracy of devices with similar appearance features or port appearance features and sub-components in the devices (such as ports, card boards, etc.), realizes accurate and efficient distinction between devices and between sub-components in devices, and can stably locate and recognize under low computing power inspection conditions.

[0095] In the embodiment of the present invention, optionally, based on a predetermined positioning and recognition model, the target image is analyzed to obtain the positioning and recognition result of each device in the computer room inspection scene, including:

[0096] The backbone network based on the predetermined positioning recognition model extracts the spatial hierarchical information of the target image, and analyzes the spatial hierarchical information of the target image to obtain multiple feature maps of the target image, and the size of each feature map in the target image is different;

[0097] The neck network based on the positioning recognition model performs a fusion operation on multiple feature maps of the target image to obtain a fusion result corresponding to the target image;

[0098] According to the head network of the positioning and recognition model, the fusion result corresponding to the target image and multiple feature maps of the target image are analyzed to obtain the positioning and recognition results of each device in the computer room inspection scene.

[0099] In an embodiment of the present invention, optionally, the backbone network is composed of multiple convolutional layers and downsampling layers to accurately capture the spatial hierarchical information in the target image and generate a semantically rich feature map.

[0100] In the embodiment of the present invention, the neck network can optionally be composed of structures such as FPN (Feature Pyramid Network) or PANet (Path Aggregation Network) to fuse feature maps of different sizes and levels, so that the positioning recognition model has better detection effects on targets of various sizes. After obtaining multiple feature maps of the target image, a pyramid is constructed based on the neck network for the multiple feature maps to enhance the multi-scale detection capability in the target image.

[0101] It can be seen that the embodiments of the present invention can also analyze the collected images through the backbone network, the neck network and the head network in sequence, obtain feature maps of different spatial levels, the fusion effect of feature maps of different spatial levels and the positioning and recognition effect, thereby improving the positioning and recognition efficiency and accuracy of the device; and when designing the backbone network, both efficiency and performance are taken into consideration to ensure that a high detection accuracy can be maintained while reducing the number of parameters; and through the neck network, the detection accuracy and overall robustness of small-sized targets can be improved; and through the head network, while maintaining a fast analysis speed, accurate position and category detection results are provided.

[0102] In the embodiment of the present invention, further optionally, the fusion result of the target image and the multiple feature maps of the target image are analyzed according to the head network of the positioning recognition model to obtain the positioning recognition result of each device in the machine room inspection scene, including:

[0103] Perform prediction box analysis on the fusion result of the target image according to the head network of the positioning recognition model to obtain multiple device prediction boxes corresponding to the target image and the confidence of each device prediction box;

[0104] For any device prediction box, the confidence of the device prediction box is compared with the preset confidence to obtain a confidence comparison result;

[0105] According to the confidence comparison result corresponding to each device prediction frame, all target device prediction frames whose confidence is greater than or equal to the preset confidence are screened out from all device prediction frames;

[0106] According to the head network of the positioning recognition model, using multiple feature maps of the target image, performing category analysis on all target device prediction boxes to obtain a category result of each target device prediction box, wherein the category result of each target device prediction box includes the probability of at least one category;

[0107] For any target device prediction box, the category with the largest probability is selected from the probabilities of all categories contained in the category results of the target device prediction box as the device category to which the target device prediction box belongs, and the actual position corresponding to the position of the target device prediction box in the target image in the computer room inspection scene is determined as the actual position corresponding to the device corresponding to the device category to which the target device prediction box belongs in the computer room inspection scene.

[0108] In an embodiment of the present invention, optionally, when multiple target device prediction frames having a position similarity greater than or equal to a preset position similarity and / or a category similarity greater than or equal to a preset category similarity are detected at the same position, the confidence of each target device prediction frame can also be determined, and the device position and device category corresponding to the target device prediction frame with a larger confidence are selected as the position, that is, the device category and device position of the corresponding device, so as to further improve the detection accuracy of the device category and device position.

[0109] It can be seen that the embodiments of the present invention can also perform prediction box, prediction box confidence and category result analysis on the feature fusion result of the image based on the head network, so as to achieve accurate distinction between the foreground and background of the collected image, and then perform accurate positioning analysis of the device position and device category, which can achieve accurate and reliable positioning and identification of the device category and device position.

[0110] In an optional embodiment, the method may further include the following steps:

[0111] Acquire first device image training data in a sample computer room inspection scene, wherein the first device image training data includes a position and a device category of a device in each of a plurality of first sample device images in the sample computer room inspection scene;

[0112] Performing a training operation on a YOLOv7 base positioning model using the first device image training data;

[0113] In the process of training the YOLOv7 basic positioning model, based on the predetermined loss network, the loss parameter of the YOLOv7 basic positioning model in the training process is calculated;

[0114] According to the loss parameters corresponding to the YOLOv7 basic positioning model, an optimization operation is performed on the YOLOv7 basic positioning model in the training process to obtain an optimized YOLOv7 basic positioning model;

[0115] The optimized YOLOv7 basic positioning model is determined as the positioning recognition model.

[0116] In this optional embodiment, optionally, different loss networks correspond to different loss parameters. Optionally, the loss network may include a confidence loss network and a classification loss network. Among them, the loss parameter corresponding to the confidence loss network is confidence, which is used to evaluate whether the prediction box predicted by the positioning recognition model contains a device, wherein the greater the confidence, the greater the probability of containing the device. Optionally, the loss parameter corresponding to the classification loss network is category probability, which is used to evaluate the difference between the device category predicted by the positioning recognition model and the actual device category, which can be evaluated by category probability, wherein the larger the predicted category probability, that is, the smaller the difference between the actual device category and the predicted category, the closer it is to the actual device category. Optionally, the classification loss network can adopt a cross entropy loss network to minimize the classification error, so that the positioning recognition model can identify categories more accurately.

[0117] In this optional embodiment, further optionally, the calculation formula for the loss parameter corresponding to the cross entropy loss network is as follows:

[0118] L n =-w n [y n ·logS(x n )+(1-y n )·log(1-S(x n ))];

[0119] Where, L n Indicates the category loss value of the YOLOv7 basic positioning model when training, W n Represents the mean width of all candidate boxes that may contain the device in the first sample device image, y n represents the true category probability corresponding to the device in the first sample device image, S(x n ) represents the predicted category probability corresponding to the predicted box of the device in the first sample device image.

[0120] In this optional embodiment, further optionally, the loss network may also include a position loss network. The loss parameter corresponding to the position loss network is the position loss, and an intersection-over-union loss network may be used, which not only considers the degree of overlap (IoU) between the predicted box and the true box, but also uses aspect ratio consistency and overlapping area factors to better handle devices of different shapes and sizes, so that the positioning recognition model is more accurate when performing position regression, that is, more accurate when performing device position positioning.

[0121] In this optional embodiment, further optionally, the calculation formula for the loss parameter corresponding to the intersection-over-intersection loss network is as follows:

[0122]

[0123] Where CIoU represents the position loss value of the YOLOv7 basic positioning model in the current training, IoU represents the intersection area between the predicted box of the device in the first sample device image and the actual box of the device, b represents the center point of the predicted box, and b gt represents the center point of the actual box, ρ 2 (b,b gt ) represents the distance between the center point of the predicted box and the actual box, c represents the diagonal length of the minimum enclosing rectangular box of the two boxes, w and h represent the width and height of the predicted box, w gt 、h gt Indicates the width and height of the actual box; a is used to evaluate whether the aspect ratio between the predicted box and the actual box is consistent, while v increases the priority of the overlapping area factor in the regression process.

[0124] In this optional embodiment, it should be noted that training the YOLOv7 basic positioning model and optimizing the YOLOv7 basic positioning model based on the corresponding loss parameter is a dynamic process. Optionally, the YOLOv7 basic positioning model after the loss parameter optimization corresponding to the preset training number based on the training number iteration can be determined as the positioning recognition model, and / or the YOLOv7 basic positioning model after the loss parameter optimization based on a value less than or equal to the preset loss value can be determined as the positioning recognition model.

[0125] In this optional embodiment, optionally, the sample computer room inspection scenario may be the computer room inspection scenario referred to above, or may be other computer room inspection scenarios.

[0126] In this optional embodiment, optionally, in order to reduce the model complexity and computing power requirements of the positioning and recognition model, multiple processing layers in the YOLOv7 backbone network can be simplified or eliminated before training, so that the trained positioning and recognition model is more compact and has a faster analysis speed, and is more suitable for computer room inspection scenarios.

[0127] It can be seen that this optional embodiment also performs training of the YOLOv7 basic positioning model and calculation of loss parameters based on the equipment images of the computer room inspection scene, which is conducive to obtaining a more applicable, accurate and reliable positioning and recognition model, thereby facilitating improving the positioning and recognition accuracy of equipment positions and categories in computer room inspection scenes.

[0128] In another optional embodiment, the method may further include the following steps:

[0129] Acquire second device image training data in a sample computer room inspection scene, wherein the second device image training data includes a position and a category corresponding to each second device sample image in a plurality of second sample device images, each second device sample image includes at least one second sample device, each second sample device includes at least three sample sub-components, the position corresponding to each second device sample image includes the device positions of all second sample devices and the positions of the sub-components included in each second sample device, and the category corresponding to each second device sample image includes the component category of the sub-components included in each second sample device;

[0130] Based on the second device image training data, a training operation is performed on the predetermined SVM basic classification recognition model until a model optimization parameter that satisfies the predetermined model optimization condition is obtained;

[0131] The model optimization parameters that meet the model optimization conditions are updated to the SVM basic classification recognition model trained this time, and the updated SVM basic classification recognition model is determined as the aforementioned classification recognition model.

[0132] In this optional embodiment, optionally, the second sample device corresponding to the second device image training data and the sample sub-components of each second sample device include multiple categories to obtain a more accurate, reliable and adaptable classification and recognition model.

[0133] In this optional embodiment, for the description of the subcomponents, please refer to the description of the relevant content in the embodiment of the present invention, which will not be repeated here.

[0134] It can be seen that this optional embodiment trains the SVM basic classification and recognition model through massive equipment image training data until the trained model optimization parameters meet the conditions, and finally configures the parameters of the SVM basic classification and recognition model trained this time with the model optimization parameters that meet the conditions, so as to obtain an accurate and reliable classification and recognition model, which is beneficial to improve the classification and position recognition accuracy of the classification and recognition model for sub-components.

[0135] In this optional embodiment, optionally, based on the second device image training data, a training operation is performed on a predetermined SVM basic classification recognition model until a model optimization parameter that satisfies a predetermined model optimization condition is obtained, including:

[0136] According to the sample computer room inspection scenario, a preset number of target model parameters are screened from the model parameters of the predetermined gray wolf optimization model, wherein, for any second sample device image, the number of all sample subcomponents in the second sample device image is greater than or equal to the preset number, and the target model parameters corresponding to the second sample device image are not repeated;

[0137] For any second sample device image in the second device image training data, based on the position of the sample subcomponent contained in the second sample device image and the category of the sample subcomponent, all target model parameters of the gray wolf optimization model are analyzed to obtain data of candidate points matching the second sample device image, where the data of the candidate points corresponding to the second sample device image include the position of the candidate point and the distance between the candidate point and each sample subcomponent in the second sample device image;

[0138] Based on the target data corresponding to the second device image training data, the predetermined SVM basic classification recognition model is trained until a model optimization parameter that satisfies the predetermined model optimization condition is obtained;

[0139] The target data corresponding to the second device image training data includes the position of the sample subcomponent contained in each second sample device image, the component category of the sample subcomponent, the position of the candidate point, the distance between the candidate point and each sample subcomponent, and the device position.

[0140] In this optional embodiment, different sample room inspection scenarios may correspond to different preset numbers, such as 3 or 5. Specifically, the greater the density of the sample subcomponents, and / or the greater the number of categories of the sample subcomponents, and / or the smaller the area of ​​the sample subcomponents, and / or the more complex the shape of the sample subcomponents, the greater the corresponding preset number.

[0141] In this optional embodiment, the number of model optimization parameters can be 2, 3, or other numbers. Further optionally, the model optimization parameters include a first model optimization parameter, a second model optimization parameter, and a third model optimization parameter. Optionally, the value ranges of the first model optimization parameter, the second model optimization parameter, and the third model optimization parameter are different from each other, or the value ranges of the second model optimization parameter and the third model optimization parameter are the same, but different from the first model optimization parameter.

[0142] In this optional embodiment, optionally, before training the SVM basic classification recognition model, corresponding first random parameters and second random parameters may be defined for each sample subcomponent to participate in the determination of candidate points. Optionally, the first random parameter and the second random parameter corresponding to each sample subcomponent may be the same or different. Specifically, for any sample subcomponent, the position of the intermediate candidate point is obtained based on the position of the sample subcomponent, the corresponding first random parameter and the second random parameter, and the position of the previous candidate point, and the position of the candidate point is determined based on the position of the intermediate candidate point corresponding to each sample subcomponent, and the category of the candidate point is determined based on the category corresponding to each sample subcomponent, so that each sample subcomponent can adjust its position according to the position of the candidate point, thereby obtaining the optimal candidate point, thereby effectively training the SVM basic classification recognition model.

[0143] It should be noted that, optionally, when it is the first iteration, the last candidate point can be 0 or a random point. Optionally, the position of the candidate point can be understood as the vector direction of the recognition center of the position of each sample sub-component.

[0144] In this optional embodiment, optionally, the model optimization conditions include that the number of training iterations is greater than or equal to a preset number of iterations and / or the model optimization parameters obtained after each training are within a predetermined parameter range; wherein the parameter range is determined based on one or more of the density of equipment in the sample computer room inspection scenario, the type of equipment, the density of sample sub-components in the equipment, the size of the sample sub-components, and the shape of the sample sub-components.

[0145] It can be seen that this optional embodiment analyzes the second device image training data through the Gray Wolf Optimization Model, and then combines the analyzed data with the second device image training data to train the SVM basic classification and recognition model to fit the optimal parameters of the SVM basic classification and recognition model, and configures the optimal parameters of the SVM basic classification and recognition model to ensure that the classification and recognition model achieves optimal performance in the category classification and position recognition tasks of the sub-components of the device, which not only improves the recognition accuracy of the classification and recognition model, but also enhances its stability and reliability in practical applications, which is beneficial to the category classification and accurate classification and recognition of the positions of the sub-components of the device.

[0146] In the embodiment of the present invention, optionally, Figure 2 It is a schematic diagram of intelligent identification of a device disclosed in the prior art. Figure 3 FIG. 1 is a schematic diagram of an embodiment of the present invention for applying the present invention to intelligently identify a device. Figure 2 As shown in the figure, the location and category of the sub-components on the equipment are obtained by manually checking the sub-components on the equipment, which is inefficient and prone to errors, especially for sub-components with very similar appearance features. Figure 3As shown, the present invention can be applied to small inspection components with similar appearance features on the equipment, such as ports and labels, to achieve effective target positioning and classification identification, and to automatically inspect multiple similar sub-components, such as whole machine information, including manufacturer, model, capacity and line rate; and port information, including optical power (input), optical power (output), service name, source, destination, alarm, optical cable relay segment 1 and optical cable relay segment 2. Figure 3 It can be seen that after applying the present invention, it can be seen that there are a total of 6 boards on the upper layer of the device, each board has 4 ports, and the used port of the 6th board can be accurately located and detected to distinguish the other 5 boards and ports.

[0147] Embodiment 2

[0148] See also Figure 4 , Figure 4 : is a schematic diagram of a structure of a device for intelligently identifying a device disclosed in an embodiment of the present invention. Figure 4 The device for intelligently identifying a device described above can be applied to any scenario where it is necessary to identify the location and category of a device and its sub-components, such as a computer room inspection scenario. Figure 4 As shown, the device for intelligently identifying a device may include:

[0149] The acquisition module 201 is used to acquire the target image in the inspection scene of the computer room. There are multiple devices in the inspection scene of the computer room, and each device includes multiple sub-components;

[0150] The analysis module 202 is used to analyze the target image based on a predetermined positioning recognition model to obtain the positioning recognition result of each device in the computer room inspection scene, wherein the positioning recognition result of each device in the computer room inspection scene includes the actual position of the device in the computer room inspection scene and the device category of the device, wherein the positioning recognition model is obtained by training based on the YOLOv7 basic positioning model;

[0151] The analysis module 202 is also used to analyze the positioning recognition results and target images of each device in the computer room inspection scene based on a predetermined classification recognition model to obtain a category recognition result of the computer room inspection scene, wherein the classification recognition model is trained based on the SVM basic classification recognition model; wherein the category recognition result of the computer room inspection scene includes the component category of the sub-components in each device in the computer room inspection scene and the position of each sub-component.

[0152] It can be seen that implementation Figure 4The described device collects a target image in a computer room inspection scene, where there are multiple devices in the computer room inspection scene, and each device includes multiple sub-components; based on a predetermined positioning and recognition model, the target image is analyzed to obtain a positioning and recognition result of each device in the computer room inspection scene, and the positioning and recognition result of each device in the computer room inspection scene includes the actual position of the device in the computer room inspection scene and the device category of the device, wherein the positioning and recognition model is obtained by training based on the YOLOv7 basic positioning model; based on a predetermined classification and recognition model, the positioning and recognition result of each device in the computer room inspection scene and the target image are analyzed to obtain a category recognition result of the computer room inspection scene, wherein the classification and recognition model is obtained by training based on the SVM basic classification and recognition model; wherein the category recognition result of the computer room inspection scene includes the component category of the sub-component in each device in the computer room inspection scene and the position of each sub-component, thereby improving the positioning and recognition accuracy of devices with similar appearance features or port appearance features and sub-components in the devices (such as ports, card boards, etc.), realizing accurate and efficient distinction between devices and between sub-components in devices, and can stably locate and recognize under low computing power inspection conditions.

[0153] In the embodiment of the present invention, optionally, the analysis module 202 analyzes the target image based on a predetermined positioning and recognition model to obtain the positioning and recognition result of each device in the computer room inspection scene in a specific manner including:

[0154] The backbone network based on the predetermined positioning recognition model extracts the spatial hierarchical information of the target image, and analyzes the spatial hierarchical information of the target image to obtain multiple feature maps of the target image, and the size of each feature map in the target image is different;

[0155] The neck network based on the positioning recognition model performs a fusion operation on multiple feature maps of the target image to obtain a fusion result corresponding to the target image;

[0156] According to the head network of the positioning and recognition model, the fusion result corresponding to the target image and multiple feature maps of the target image are analyzed to obtain the positioning and recognition results of each device in the computer room inspection scene.

[0157] In an embodiment of the present invention, optionally, the backbone network is composed of multiple convolutional layers and downsampling layers to accurately capture the spatial hierarchical information in the target image and generate a semantically rich feature map.

[0158] In the embodiment of the present invention, the neck network can optionally be composed of structures such as FPN (Feature Pyramid Network) or PANet (Path Aggregation Network) to fuse feature maps of different sizes and levels, so that the positioning recognition model has better detection effects on targets of various sizes. After obtaining multiple feature maps of the target image, a pyramid is constructed based on the neck network for the multiple feature maps to enhance the multi-scale detection capability in the target image.

[0159] It can be seen that implementation Figure 4 The described device can also analyze the collected images through the backbone network, neck network and head network in sequence to obtain feature maps of different spatial levels, fusion effects of feature maps of different spatial levels and positioning and recognition effects, thereby improving the positioning and recognition efficiency and accuracy of the equipment; and when designing the backbone network, both efficiency and performance are taken into consideration to ensure that high detection accuracy can be maintained while reducing the number of parameters; and through the neck network, the detection accuracy and overall robustness of small-sized targets can be improved; and through the head network, while maintaining a fast analysis speed, accurate position and category detection results are provided.

[0160] In the embodiment of the present invention, optionally, the analysis module 202 analyzes the fusion result of the target image and the multiple feature maps of the target image according to the head network of the positioning recognition model to obtain the specific method of the positioning recognition result of each device in the computer room inspection scene, including:

[0161] Perform prediction box analysis on the fusion result of the target image according to the head network of the positioning recognition model to obtain multiple device prediction boxes corresponding to the target image and the confidence of each device prediction box;

[0162] For any device prediction box, the confidence of the device prediction box is compared with the preset confidence to obtain a confidence comparison result;

[0163] According to the confidence comparison result corresponding to each device prediction frame, all target device prediction frames whose confidence is greater than or equal to the preset confidence are screened out from all device prediction frames;

[0164] According to the head network of the positioning recognition model, using multiple feature maps of the target image, performing category analysis on all target device prediction boxes to obtain a category result of each target device prediction box, wherein the category result of each target device prediction box includes the probability of at least one category;

[0165] For any target device prediction box, the category with the largest probability is selected from the probabilities of all categories contained in the category results of the target device prediction box as the device category to which the target device prediction box belongs, and the actual position corresponding to the position of the target device prediction box in the target image in the computer room inspection scene is determined as the actual position corresponding to the device corresponding to the device category to which the target device prediction box belongs in the computer room inspection scene.

[0166] In an embodiment of the present invention, optionally, when multiple target device prediction frames having a position similarity greater than or equal to a preset position similarity and / or a category similarity greater than or equal to a preset category similarity are detected at the same position, the confidence of each target device prediction frame can also be determined, and the device position and device category corresponding to the target device prediction frame with a larger confidence are selected as the position, that is, the device category and device position of the corresponding device, so as to further improve the detection accuracy of the device category and device position.

[0167] It can be seen that implementation Figure 4 The described device can also perform prediction box, prediction box confidence and category result analysis on the feature fusion results of the image based on the head network, so as to achieve accurate distinction between the foreground and background of the collected image, and then perform accurate positioning analysis of the device position and device category, which can achieve accurate and reliable positioning and identification of the device category and device position.

[0168] In an optional embodiment, if Figure 5 As shown, Figure 5 is a schematic diagram of the structure of another device for intelligently identifying a device disclosed in an embodiment of the present invention, such as Figure 5 As shown, the device also includes:

[0169] A first acquisition module 203 is used to acquire first device image training data in a sample computer room inspection scene, wherein the first device image training data includes a position and a device category of a device in each of a plurality of first sample device images in the sample computer room inspection scene;

[0170] A first training module 204 is used to perform a training operation on a YOLOv7 basic positioning model using first device image training data;

[0171] The calculation module 205 is used to calculate the loss parameters of the YOLOv7 basic positioning model in the training process based on the predetermined loss network during the training of the YOLOv7 basic positioning model;

[0172] The optimization module 206 is used to perform an optimization operation on the YOLOv7 basic positioning model in the training process according to the loss parameter corresponding to the YOLOv7 basic positioning model to obtain an optimized YOLOv7 basic positioning model;

[0173] The first determination module 207 is used to determine the optimized YOLOv7 basic positioning model, which is a positioning recognition model.

[0174] In this optional embodiment, optionally, different loss networks correspond to different loss parameters. Optionally, the loss network may include a confidence loss network and a classification loss network. Among them, the loss parameter corresponding to the confidence loss network is confidence, which is used to evaluate whether the prediction box predicted by the positioning recognition model contains a device, wherein the greater the confidence, the greater the probability of containing the device. Optionally, the loss parameter corresponding to the classification loss network is category probability, which is used to evaluate the difference between the device category predicted by the positioning recognition model and the actual device category, which can be evaluated by category probability, wherein the larger the predicted category probability, that is, the smaller the difference between the actual device category and the predicted category, the closer it is to the actual device category. Optionally, the classification loss network can adopt a cross entropy loss network to minimize the classification error, so that the positioning recognition model can identify categories more accurately.

[0175] In this optional embodiment, further optionally, the loss network may also include a position loss network. The loss parameter corresponding to the position loss network is the position loss, and an intersection-over-union loss network may be used, which not only considers the degree of overlap (IoU) between the predicted box and the true box, but also uses aspect ratio consistency and overlapping area factors to better handle devices of different shapes and sizes, so that the positioning recognition model is more accurate when performing position regression, that is, more accurate when performing device position positioning.

[0176] In this optional embodiment, it should be noted that training the YOLOv7 basic positioning model and optimizing the YOLOv7 basic positioning model based on the corresponding loss parameter is a dynamic process. Optionally, the YOLOv7 basic positioning model after the loss parameter optimization corresponding to the preset training number based on the training number iteration can be determined as the positioning recognition model, and / or the YOLOv7 basic positioning model after the loss parameter optimization based on a value less than or equal to the preset loss value can be determined as the positioning recognition model.

[0177] In this optional embodiment, optionally, the sample computer room inspection scenario may be the computer room inspection scenario referred to above, or may be other computer room inspection scenarios.

[0178] It can be seen that implementation Figure 5 The described device can also train the YOLOv7 basic positioning model and calculate the loss parameters based on the equipment images of the computer room inspection scene, which is conducive to obtaining a more applicable, accurate and reliable positioning and recognition model, thereby facilitating improving the positioning and recognition accuracy of the equipment position and category in the computer room inspection scene.

[0179] In another optional embodiment, Figure 5 As shown, the device may also include:

[0180] The second acquisition module 208 is used to acquire second device image training data in the sample computer room inspection scene, wherein the second device image training data includes a position and a category corresponding to each second device sample image in a plurality of second sample device images, each second device sample image includes at least one second sample device, each second sample device includes at least three sample sub-components, the position corresponding to each second device sample image includes the device positions of all second sample devices and the positions of the sub-components included in each second sample device, and the category corresponding to each second device sample image includes the component category of the sub-components included in each second sample device;

[0181] The second training module 209 is used to perform a training operation on the predetermined SVM basic classification recognition model based on the second device image training data until a model optimization parameter that meets the predetermined model optimization condition is obtained;

[0182] The second determination module 210 is used to update the model optimization parameters that meet the model optimization conditions to the SVM basic classification recognition model trained this time, and determine the updated SVM basic classification recognition model as the aforementioned classification recognition model.

[0183] In this optional embodiment, optionally, the second sample device corresponding to the second device image training data and the sample sub-components of each second sample device include multiple categories to obtain a more accurate, reliable and adaptable classification and recognition model.

[0184] In this optional embodiment, for the description of the subcomponents, please refer to the description of the relevant content in the embodiment of the present invention, which will not be repeated here.

[0185] It can be seen that implementation Figure 5 The described device can train the SVM basic classification and recognition model through massive equipment image training data until the trained model optimization parameters meet the conditions. Finally, the model optimization parameters that meet the conditions are used to configure the parameters of the SVM basic classification and recognition model trained this time, so as to obtain an accurate and reliable classification and recognition model, which is beneficial to improve the classification and position recognition accuracy of the classification and recognition model for sub-components.

[0186] In this optional embodiment, optionally, the second training module 209 performs a training operation on a predetermined SVM basic classification recognition model based on the second device image training data until a model optimization parameter that satisfies a predetermined model optimization condition is obtained. Specifically, the method includes:

[0187] According to the sample computer room inspection scenario, a preset number of target model parameters are screened from the model parameters of the predetermined gray wolf optimization model, wherein, for any second sample device image, the number of all sample subcomponents in the second sample device image is greater than or equal to the preset number, and the target model parameters corresponding to the second sample device image are not repeated;

[0188] For any second sample device image in the second device image training data, based on the position of the sample subcomponent contained in the second sample device image and the category of the sample subcomponent, all target model parameters of the gray wolf optimization model are analyzed to obtain data of candidate points matching the second sample device image, where the data of the candidate points corresponding to the second sample device image include the position of the candidate point and the distance between the candidate point and each sample subcomponent in the second sample device image;

[0189] Based on the target data corresponding to the second device image training data, the predetermined SVM basic classification recognition model is trained until a model optimization parameter that satisfies the predetermined model optimization condition is obtained;

[0190] The target data corresponding to the second device image training data includes the position of the sample subcomponent contained in each second sample device image, the component category of the sample subcomponent, the position of the candidate point, the distance between the candidate point and each sample subcomponent, and the device position.

[0191] In this optional embodiment, different sample room inspection scenarios may correspond to different preset numbers, such as 3 or 5. Specifically, the greater the density of the sample subcomponents, and / or the greater the number of categories of the sample subcomponents, and / or the smaller the area of ​​the sample subcomponents, and / or the more complex the shape of the sample subcomponents, the greater the corresponding preset number.

[0192] In this optional embodiment, the number of model optimization parameters can be 2, 3, or other numbers. Further optionally, the model optimization parameters include a first model optimization parameter, a second model optimization parameter, and a third model optimization parameter. Optionally, the value ranges of the first model optimization parameter, the second model optimization parameter, and the third model optimization parameter are different from each other, or the value ranges of the second model optimization parameter and the third model optimization parameter are the same, but different from the first model optimization parameter.

[0193] In this optional embodiment, optionally, before training the SVM basic classification recognition model, corresponding first random parameters and second random parameters may be defined for each sample subcomponent to participate in the determination of candidate points. Optionally, the first random parameter and the second random parameter corresponding to each sample subcomponent may be the same or different. Specifically, for any sample subcomponent, the position of the intermediate candidate point is obtained based on the position of the sample subcomponent, the corresponding first random parameter and the second random parameter, and the position of the previous candidate point, and the position of the candidate point is determined based on the position of the intermediate candidate point corresponding to each sample subcomponent, and the category of the candidate point is determined based on the category corresponding to each sample subcomponent, so that each sample subcomponent can adjust its position according to the position of the candidate point, thereby obtaining the optimal candidate point, thereby effectively training the SVM basic classification recognition model.

[0194] It should be noted that, optionally, when it is the first iteration, the last candidate point can be 0 or a random point. Optionally, the position of the candidate point can be understood as the vector direction of the recognition center of the position of each sample sub-component.

[0195] In this optional embodiment, optionally, the model optimization conditions include that the number of training iterations is greater than or equal to a preset number of iterations and / or the model optimization parameters obtained after each training are within a predetermined parameter range; wherein the parameter range is determined based on one or more of the density of equipment in the sample computer room inspection scenario, the type of equipment, the density of sample sub-components in the equipment, the size of the sample sub-components, and the shape of the sample sub-components.

[0196] It can be seen that implementation Figure 5 The described device can analyze the second device image training data through the gray wolf optimization model, and then combine the analyzed data with the second device image training data to train the SVM basic classification and recognition model to fit the optimal parameters of the SVM basic classification and recognition model, and configure the optimal parameters of the SVM basic classification and recognition model to ensure that the classification and recognition model achieves optimal performance in the category classification and position recognition tasks of the sub-components of the equipment, which not only improves the recognition accuracy of the classification and recognition model, but also enhances its stability and reliability in practical applications, which is beneficial to the category classification and accurate classification and recognition of the positions of the sub-components of the equipment.

[0197] Embodiment 3

[0198] See also Figure 6 , Figure 6 is a schematic diagram of the structure of another device for intelligently identifying a device disclosed in an embodiment of the present invention. The device can be applied to any scenario where it is necessary to identify the location category of a device and its subcomponents, such as a computer room inspection scenario. Figure 6As shown, the device for intelligently identifying a device may include:

[0199] A memory 301 storing executable program codes;

[0200] a processor 302 coupled to the memory 301;

[0201] The processor 302 calls the executable program code stored in the memory 301 to execute the steps of the method for intelligently identifying a device described in the first embodiment of the present invention.

[0202] Embodiment 4

[0203] An embodiment of the present invention discloses a computer-readable storage medium, which stores computer instructions. When the computer instructions are called, they are used to execute the steps of the method for intelligently identifying a device described in the first embodiment of the present invention.

[0204] Embodiment 5

[0205] An embodiment of the present invention discloses a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program, and the computer program is operable to enable a computer to execute the steps of the method for intelligently identifying a device described in Embodiment 1.

[0206] The device embodiments described above are only illustrative, wherein the modules described as separate components may or may not be physically separated, and the components displayed as modules may or may not be physical modules, i.e., they may be located in one place, or they may be distributed on multiple network modules. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Those of ordinary skill in the art may understand and implement it without creative work.

[0207] Through the specific description of the above embodiments, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on such an understanding, the above technical solution can be essentially or partly contributed to the prior art in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, and the storage medium includes a read-only memory (ROM), a random access memory (RAM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), a one-time programmable read-only memory (OTPROM), an electronically erasable rewritable read-only memory (EEPROM), a compact disc (CD-ROM) or other optical disc storage, magnetic disk storage, magnetic tape storage, or any other computer-readable medium that can be used to carry or store data.

[0208] Finally, it should be noted that the method and apparatus for intelligently identifying a device disclosed in the embodiments of the present invention disclose only the preferred embodiments of the present invention, and are only used to illustrate the technical solution of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligently identifying a device, characterized in that: The method is applied to a computer room inspection scenario, and the method includes: Collecting a target image in the computer room inspection scene, where there are multiple devices in the computer room inspection scene, and each of the devices includes multiple sub-components; Based on a predetermined positioning recognition model, the target image is analyzed to obtain a positioning recognition result of each of the devices in the computer room inspection scene, wherein the positioning recognition result of each of the devices in the computer room inspection scene includes the actual position of the device in the computer room inspection scene and the device category of the device, wherein the positioning recognition model is obtained by training based on the YOLOv7 basic positioning model; Based on a predetermined classification and recognition model, the positioning and recognition results of each device in the computer room inspection scene and the target image are analyzed to obtain a category recognition result of the computer room inspection scene, wherein the classification and recognition model is obtained by training based on the SVM basic classification and recognition model; The category identification result of the computer room inspection scene includes the component category and the position of the subcomponent of each subcomponent in the equipment in the computer room inspection scene.

2. The method for intelligently identifying a device according to claim 1, characterized in that: The target image is analyzed based on a predetermined positioning and recognition model to obtain a positioning and recognition result of each device in the machine room inspection scene, including: Extracting spatial hierarchical information of the target image based on a backbone network of a predetermined positioning recognition model, and analyzing the spatial hierarchical information of the target image to obtain a plurality of feature maps of the target image, wherein the size of each feature map in the target image is different; Based on the neck network of the positioning recognition model, a fusion operation is performed on multiple feature maps of the target image to obtain a fusion result corresponding to the target image; The fusion result corresponding to the target image and multiple feature maps of the target image are analyzed according to the head network of the positioning and recognition model to obtain the positioning and recognition results of each of the devices in the computer room inspection scene.

3. The method for intelligently identifying a device according to claim 2, characterized in that: The head network according to the positioning recognition model analyzes the fusion result of the target image and the multiple feature maps of the target image to obtain the positioning recognition result of each device in the machine room inspection scene, including: Performing prediction frame analysis on the fusion result of the target image according to the head network of the positioning recognition model to obtain multiple device prediction frames corresponding to the target image and the confidence of each of the device prediction frames; For any of the device prediction frames, comparing the confidence of the device prediction frame with a preset confidence to obtain a confidence comparison result; According to the confidence comparison result corresponding to each of the device prediction frames, all target device prediction frames whose confidence is greater than or equal to the preset confidence are screened out from all the device prediction frames; According to the head network of the positioning recognition model, using multiple feature maps of the target image, performing category analysis on all the target device prediction boxes to obtain a category result of each of the target device prediction boxes, wherein the category result of each of the target device prediction boxes includes a probability of at least one category; For any of the target device prediction boxes, the category with the largest probability is selected from the probabilities of all categories contained in the category results of the target device prediction box as the device category to which the target device prediction box belongs, and the actual position corresponding to the position of the target device prediction box in the target image in the computer room inspection scene is determined to be the actual position corresponding to the device corresponding to the device category to which the target device prediction box belongs in the computer room inspection scene.

4. The method for intelligently identifying a device according to any one of claims 1 to 3, characterized in that: The method further comprises: Acquire first device image training data in a sample computer room inspection scene, wherein the first device image training data includes a position and a device category of a device in each of a plurality of first sample device images in the sample computer room inspection scene; Performing a training operation on the YOLOv7 basic positioning model using the first device image training data; In the process of training the YOLOv7 basic positioning model, based on the predetermined loss network, calculating the loss parameter of the YOLOv7 basic positioning model in the training process; According to the loss parameter corresponding to the YOLOv7 basic positioning model, an optimization operation is performed on the YOLOv7 basic positioning model in the training process to obtain the optimized YOLOv7 basic positioning model; The optimized YOLOv7 basic positioning model is determined as the positioning recognition model.

5. The method for intelligently identifying a device according to claims 1-3, characterized in that: The method further comprises: Acquire second device image training data in a sample computer room inspection scene, wherein the second device image training data includes a position and a category corresponding to each second device sample image in a plurality of second sample device images, each second device sample image includes at least one second sample device, each second sample device includes at least three sample sub-components, the position corresponding to each second device sample image includes the device positions of all the second sample devices and the positions of the sub-components included in each second sample device, and the category corresponding to each second device sample image includes the component category of the sub-components included in each second sample device; Based on the second device image training data, a training operation is performed on the predetermined SVM basic classification recognition model until a model optimization parameter that satisfies the predetermined model optimization condition is obtained; The model optimization parameters that meet the model optimization conditions are updated to the SVM basic classification recognition model of this training, and the updated SVM basic classification recognition model is determined as the classification recognition model.

6. The method for intelligently identifying a device according to claim 5, characterized in that: The performing of a training operation on the predetermined SVM basic classification recognition model based on the second device image training data until obtaining a model optimization parameter that satisfies a predetermined model optimization condition includes: According to the sample computer room inspection scenario, a preset number of target model parameters are screened from the model parameters of the predetermined gray wolf optimization model, wherein, for any second sample device image, the number of all the sample subcomponents in the second sample device image is greater than or equal to the preset number, and the target model parameters corresponding to the second sample device image are not repeated; For any second sample device image in the second device image training data, based on the position of the sample subcomponent contained in the second sample device image and the category of the sample subcomponent, all the target model parameters of the gray wolf optimization model are analyzed to obtain data of candidate points matching the second sample device image, wherein the data of the candidate points corresponding to the second sample device image include the position of the candidate point and the distance between the candidate point and each of the sample subcomponents in the second sample device image; Based on the target data corresponding to the second device image training data, the predetermined SVM basic classification recognition model is trained until a model optimization parameter that satisfies the predetermined model optimization condition is obtained; The target data corresponding to the second device image training data includes the position of the sample sub-component contained in each of the second sample device images, the component category of the sample sub-component, the position of the candidate point, the distance between the candidate point and each sample sub-component, and the device position.

7. The method for intelligently identifying a device according to claim 5 or 6, characterized in that: The model optimization condition includes that the number of training iterations is greater than or equal to a preset number of iterations and / or the model optimization parameters obtained after each training are within a predetermined parameter range; The parameter range is determined based on one or more of the density of equipment in the sample computer room inspection scene, the type of equipment, the density of sample sub-components in the equipment, the size of the sample sub-components, and the shape of the sample sub-components.

8. A device for intelligently identifying a device, characterized in that: The device is applied to a computer room inspection scenario, and the device includes: A collection module, used for collecting target images in the inspection scene of the computer room, wherein there are multiple devices in the inspection scene of the computer room, and each of the devices includes multiple sub-components; An analysis module is used to analyze the target image based on a predetermined positioning recognition model to obtain a positioning recognition result of each of the devices in the computer room inspection scene, wherein the positioning recognition result of each of the devices in the computer room inspection scene includes the actual position of the device in the computer room inspection scene and the device category of the device, wherein the positioning recognition model is obtained by training based on the YOLOv7 basic positioning model; The analysis module is further used to analyze the positioning recognition results of each of the devices in the computer room inspection scene and the target image based on a predetermined classification recognition model to obtain a category recognition result of the computer room inspection scene, wherein the classification recognition model is obtained by training based on the SVM basic classification recognition model; The category identification result of the computer room inspection scene includes the component category and the position of the subcomponent of each subcomponent in the equipment in the computer room inspection scene.

9. A device for intelligently identifying a device, characterized in that: The device comprises: A memory storing executable program code; a processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the method for intelligently identifying a device as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, and when the computer instructions are called, they are used to execute the method for intelligently identifying a device as described in any one of claims 1-7.

Images