Characteristic distillation tree line fault identification method based on RT-DETR algorithm

Abstract

The invention discloses a feature distillation tree line fault identification method based on an RT-DETR algorithm, and belongs to the technical field of electric power disaster prevention, and the method comprises the steps: constructing a tree line fault image data set, and dividing the tree line fault image data set into confrontation image data and annotation image data; training the confrontation image data to obtain super-resolution image data; mixing the super-resolution image data with the annotated image data, and dividing into a training set, a verification set and a test set; respectively training, verifying and testing the teacher model and the student model to obtain a performance optimal weight; utilizing the optimal performance weight to obtain a student model after feature distillation; obtaining a pruned student model by using the student model after distillation; performing distillation fine tuning to obtain a tree line fault identification model; and deploying the tree line fault identification model to an edge end system. Compared with the prior art, the method can reduce the manual detection cost, improves the real-time performance of fault detection, reduces the dependence on a cloud server, and is more suitable for a tree line fault recognition scene.

Description

Technical Field

[0001] The present invention belongs to the technical field of power disaster prevention, involving the technical fields of power engineering and deep learning image recognition, and particularly relates to a feature distillation tree-line fault recognition method based on the RT-DETR algorithm. Background Art

[0002] Transmission line tree-line faults refer to the situation where trees or other vegetation are overly close to or even in contact with high-voltage transmission lines, resulting in fault phenomena such as arc discharge and short-circuit tripping. Such faults not only cause power supply interruptions but may also cause equipment damage and even secondary disasters such as forest fires. Currently, most power line inspection methods still rely on manual inspections. Manual inspection methods have high labor intensity, long cycles, and low efficiency. The levels of line inspection personnel vary, and there may be missed inspections and misinspections, making it difficult to ensure comprehensive and accurate hidden danger investigation. Currently, there are also methods using unmanned aerial vehicles for inspection, but they have problems such as data transmission delay, high network bandwidth requirements, high energy consumption, and high requirements for servers, making it difficult to meet the real-time requirements of tree-line fault detection. Summary of the Invention

[0003] The purpose of the present invention is to provide a tree-line fault recognition method based on feature distillation, which is applicable to edge device operation, to solve the problems of insufficient real-time performance and high dependence on cloud servers in existing tree-line fault detection technologies, and at the same time reduce the cost of manual detection.

[0004] To achieve the above purpose, the present invention provides the following technical solutions:

[0005] A feature distillation tree-line fault recognition method based on the RT-DETR algorithm, and the steps of the recognition method are as follows:

[0006] S1. Establish tree-line fault image data, including adversarial image data and labeled image data;

[0007] S2. Use the adversarial image data obtained in step S1 for training, validating, and testing the SRGAN model, and use the weights with the optimal performance for super-resolution reconstruction of all images in the adversarial images to obtain super-resolution image data;

[0008] S3. Divide the labeled image data obtained in step S1 and the super-resolution image data obtained in step S2 into a training set, a validation set, and a test set, and train, validate, and test the teacher model on the divided data sets respectively. After training, validating, and testing are completed, obtain the weights with the optimal performance of the teacher model;

[0009] S4. Train, validate, and test the student model on the data sets divided in step S3. After training, validating, and testing are completed, obtain the weights with the optimal performance of the student model;

[0010] S5. Use the teacher model obtained in step S3 to perform feature distillation on the student model to obtain the student model after feature distillation;

[0011] S6. Perform pruning operation on the student model after distillation training to obtain the pruned student model;

[0012] S7. Perform distillation fine-tuning on the pruned student model to obtain the fine-tuned student model;

[0013] S8. Deploy the fine-tuned student model to the edge side, and based on the technical requirements of the edge side, perform format conversion on the model and weight files to realize the deployment of the model on the edge side.

[0014] Furthermore, in step S1, a tree-line discharge experimental platform is used to simulate the real tree-line discharge phenomenon, and an image acquisition device is used to obtain image data when the tree-line fault occurs, and it is divided into adversarial image data and partially labeled image data.

[0015] Furthermore, in step S2, in the SRGAN model, the mean squared error (MSE) loss is used to train the generator G to generate high-resolution images, and then the discriminator is trained to learn to distinguish between real images and the generated super-resolution images. The adversarial loss is used to train the generator, and the entire training process adopts an alternating training strategy until the optimal G is obtained. During the training, G adopts the perceptual loss function L MSE and L VGG , and its mathematical formula is:

[0016]

[0017] where G(I LR ) i is the i-th pixel value of the super-resolution image generated by the generator G, I HR,i is the i-th pixel value of the real high-resolution image, and N is the total number of pixels in the image;

[0018]

[0019] φ(·) is the intermediate layer feature map of the VGG pre-trained network, G(I LR ) is the super-resolution image output by the generator, I HR is the real high-resolution image, and ||·|| 2 represents the L2 norm.

[0020] Furthermore, in step S2, for super-resolution reconstruction, the results of super-resolution reconstruction are evaluated through three metrics: peak signal-to-noise ratio (PSNR), mean squared error (MSE), and structural similarity index measure (SSIM), so as to obtain super-resolution image data. The calculation formulas of PSNR, MSE, and SSIM are as follows:

[0021]

[0022] I SR (i, j) is the pixel value of the super-resolution image, I HR (i, j) is the pixel of the true high-resolution image, and M and N are the width and height of the image respectively;

[0023]

[0024] In the formula, MAX I is the maximum pixel value of the image, and MSE is the mean square error;

[0025]

[0026] Among them, μ SR and μ HR are the brightness of the super-resolution image and the true image, σ SR and σ HR are the contrast, σ SR,HR is the structural information, and C1 and C2 are stability factors;

[0027] LPIPS(I SR , I HR ) = ||F(I SR ) - F(I HR )|| 2

[0028] Among them, F(·) is the feature representation extracted by the pre-trained VGG network.

[0029] Furthermore, in step S3, the RT-DETR-X model is selected as the teacher model, and in step S4, the RT-DETR-L model is selected as the student model.

[0030] Furthermore, in step S5, improved global-local feature distillation is used, and the formula for its global distillation loss is:

[0031]

[0032] Among them, G T is the global feature vector of the teacher model, and G S is the global feature vector of the student model, and their calculation formulas are respectively:

[0033]

[0034] Among them, F T (i, j) is the feature map value of the teacher model at the (i, j) position, and H and W are the height and width of the feature map respectively;

[0035]

[0036] Among them, H and W are the height and width of the feature map respectively;

[0037] The formula for calculating its local distillation loss is:

[0038] L LFD = |||L T - L S ||| 2

[0039] Among them, L T and L S represent the regional features of the teacher model and the regional features of the student model respectively.

[0040] Furthermore, in step S6, a structured channel pruning method based on the channel scaling factor γ in the Batch Normalization layer is adopted. Specifically, for each convolutional module containing the BatchNorm layer in the student model, the scaling parameter γ corresponding to its c-th channel is extracted c , and each channel in the BatchNormalization layer contains a learnable scaling factor γ and an offset factor β, which are used to adjust the amplitude and offset of the normalized features respectively. The formula for taking the absolute value of the scaling parameter γ c as the importance score of this channel is as follows:

[0041] I c = |γ c |

[0042] Among them, γ c represents the scaling factor of the c-th channel, and I c is the importance score of this channel;

[0043] After obtaining the importance scores I1, I2,..., I c of all channels, they are sorted in ascending order. Set the pruning ratio p ∈ (0, 1), and mark the channels with the lowest scores as channels to be pruned. For the channels marked as channels to be pruned, perform the following pruning operations:

[0044] In the current convolutional layer, delete the convolutional kernel parameters corresponding to the pruned channels; in the BatchNormalization layer corresponding to this convolutional layer, delete the scaling factor γ and offset β parameters of the corresponding channels; in the subsequent connected convolutional layers, delete the weight parameters of the input channels corresponding to the pruned channels; for network architectures containing residual connections or feature fusion structures, synchronously adjust the input and output channel dimensions of the corresponding paths.

[0045] Furthermore, in step S7, distillation fine-tuning is performed, and the distillation loss calculation formula is:

[0046] L total =L task +αL feature +βL logits

[0047] Among them, α, β are hyper parameters, L task is the mission loss, L feature Feature distillation loss, L logits The calculation formulas for output distillation loss are:

[0048] L task =L cls +L reg +L obj

[0049] Among them, L cls is the classification loss, L reg is the bounding box regression loss, L obj is the confidence loss;

[0050]

[0051] in, is the feature map of the i-th layer of the student model, is the i-th layer feature map corresponding to the teacher model, and N is the number of selected distillation feature layers;

[0052]

[0053] Z T With Z S They represent the logits output of the teacher model and the student model respectively, T is the temperature parameter that controls the smoothness of softmax, and KL is the Kullback-Leibler divergence.

[0054] Furthermore, in step S8, the edge end triggers an alarm when a branch or a flame is detected, and starts uploading the real-time image of the alarm detection to the system.

[0055] The principle and beneficial effects of this technical solution:

[0056] Aiming at the problems of insufficient real-time performance and high dependence on cloud servers in the existing tree line fault detection technology, a feature distillation tree line fault identification method based on the RT-DETR algorithm is proposed. This method has the advantages of high detection accuracy and model lightweight, and is more suitable for tree line fault identification. To improve the model's ability to identify flame features in complex environments, the SRGAN model is used to preprocess the image for super-resolution reconstruction to enhance the image detail features and improve the detection accuracy of flame data. The RT-DETR algorithm is used to detect three types of targets in the image: flame, tree branch, and wire. By transferring the knowledge learned in the teacher model to the student model, the recognition ability and generalization ability of the student model are enhanced. Subsequently, pruning operations are performed on the student model to further reduce the model complexity while maintaining a high detection accuracy to meet its deployment requirements on edge devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is a schematic structural diagram of a feature distillation tree line fault identification method based on the RT-DETR algorithm. DETAILED DESCRIPTION OF THE INVENTION

[0058] The present invention will be further described in detail below in conjunction with the drawings and embodiments:

[0059] As Figure 1 shown, in order to further explain the technical solution of the present invention, the present invention will be elaborated in detail below through specific embodiments.

[0060] A feature distillation tree line fault identification method based on the RT-DETR algorithm disclosed in this embodiment has the following steps of use:

[0061] S1. Establish tree line fault image data, which includes adversarial image data and labeled image data;

[0062] In this embodiment, a real tree line discharge is simulated through a tree line discharge experiment platform, and an image acquisition device is used to collect multiple (the number should be as large as possible, 7000 images were collected in the embodiment experiment) images containing tree line discharge. Then, Gaussian blur is performed on some of the images (two thousand images were used in this example) to obtain augmented images. The images used for Gaussian blur and the images generated by using Gaussian blur are the labeled image data, and the rest are adversarial data images.

[0063] S2. Divide the labeled image data obtained in step S1 into a training set, a validation set, and a test set according to a certain ratio (7:2:1 in the embodiment experiment). The training set is used to train the SRGAN model. Set the initial learning rates of the generator and the discriminator to 1e-4, the minimum learning rate to 1e-7, the learning rate decay method to cos, use the Adam optimizer, set the adversarial loss ratio to 0.005, set the upsampling multiple to 4, and train for 500 epochs with a batch size of 16. The validation set is used to evaluate the performance of the model at each stage and save the best weights; the test set is used to independently evaluate the generalization ability of the model after training. Finally, use the model weights with the best performance on the validation set to perform super-resolution reconstruction on all low-resolution images, thereby generating super-resolution image data.

[0064] In this step, the generator G of SRGAN receives a low-resolution image and outputs a high-resolution image. The task of the discriminator D is to determine whether the input high-resolution image is a real image or a pseudo-high-resolution image reconstructed by the generator G. During training, by continuously optimizing the discriminator D and the generator G, the images generated by G become closer and closer to the distribution of real images, thereby improving the reconstruction effect and finally obtaining the generator G with the best performance. In training, G uses the perceptual loss function L MSE and L VGG , and its mathematical formula is:

[0065]

[0066] where G(I LR ) i is the i-th pixel value of the super-resolution image generated by the generator G, I HR,i is the i-th pixel value of the real high-resolution image, and N is the total number of pixels in the image;

[0067]

[0068] φ(·) is the intermediate layer feature map of the VGG pre-trained network, G(I LR ) is the super-resolution image output by the generator, I HR is the real high-resolution image, and ||·|| 2 represents the L2 norm.

[0069] After the model training is completed, the results of super-resolution reconstruction are evaluated through three metrics: PSNR (Peak Signal-to-Noise Ratio), MSE (Mean Squared Error), and SSIM (Structural Similarity Index), so as to obtain super-resolution image data. The calculation formulas for PSNR, MSE, and SSIM are as follows:

[0070]

[0071] I SR (i, j) is the pixel value of the super-resolution image, and I HR (i, j) is the pixel of the true high-resolution image. M and N are the width and height of the image respectively;

[0072]

[0073] In the formula, MAX I is the maximum pixel value of the image, and MSE is the mean squared error;

[0074]

[0075] Among them, μ SR and μ HR are the brightness of the super-resolution image and the true image, σ SR and σ HR are the contrast, σ SR,HR is the structural information, and C1 and C2 are stability factors;

[0076] LPIPS(I SR , I HR ) = ||F(I SR ) - F(I HR )|| 2

[0077] Among them, F(·) is the feature representation extracted by the pre-trained VGG network.

[0078] S3. Use an image annotation tool (LabelImg is used in the example experiment) to annotate the labeled image data obtained in step S1 and the super-resolution image data obtained in step S2 with labels such as flame, branch, and wire to obtain an image dataset.

[0079] S4. Divide the image dataset obtained in step S3 into a training set, a validation set, and a test set according to a certain ratio (7:2:1 in the example experiment). Train, validate, and test the teacher model on the divided dataset. In this example, the RT-DETR-X model is selected as the teacher model. In this example, the training parameters of the teacher model are set as follows: the initial learning rate is set to 1e-4, the optimizer uses AdamW, the learning rate decay strategy selects cos, the weight decay coefficient is 0.01, the batch size is 32, and the number of training epochs is 500. During the model training process, training, validation, and testing are carried out in sequence, and the optimal weights of the corresponding teacher model are saved when the performance on the validation set is the best.

[0080] S5. Divide the image dataset obtained in step S3 into a training set, a validation set, and a test set according to a certain ratio (7:2:1 in the example experiment). Train, validate, and test the student model on the divided dataset. In this example, the RT-DETR-L model is selected as the student model. In this example, the training parameters of the student model are set as follows: the initial learning rate is set to 5e-4, the optimizer uses AdamW, the learning rate decay strategy selects cos, the weight decay coefficient is 0.01, the batch size is 32, and the number of training epochs is 300. During the model training process, training, validation, and testing are carried out in sequence, and the optimal weights of the corresponding student model are saved when the performance on the validation set is the best.

[0081] S6. Use the optimal weights of the teacher model obtained in step S4 and the optimal weights of the student model obtained in step S5 to perform global-local feature distillation training on the student model to obtain the distilled-trained student model. The step S4 uses improved global-local feature distillation, and its global distillation loss calculation formula is:

[0082]

[0083] where, G T is the global feature vector of the teacher model, and G S is the global feature vector of the student model, and their calculation formulas are respectively:

[0084]

[0085] where, F T (i,j) is the feature map value of the teacher model at the (i,j) position, and H and W are respectively the height and width of the feature map;

[0086]

[0087] where, H and W are respectively the height and width of the feature map;

[0088] The calculation formula for its local distillation loss is as follows:

[0089] L LFD = |||L T - L S ||| 2

[0090] Where, L T and L S respectively represent the regional features of the teacher model and the regional features of the student model.

[0091] S7. Perform pruning operations on the student model after distillation training to obtain the pruned student model. Adopt a structured channel pruning method based on the channel scaling factor γ in the BatchNormalization layer. Specifically, for each convolutional module containing the BatchNorm layer in the student model, extract the scaling parameter γ corresponding to its c-th channel c , Each channel in the BatchNormalization layer contains a learnable scaling factor γ and an offset factor β, which are used to adjust the amplitude and offset of the normalized features respectively. The formula for taking the absolute value of the scaling parameter γ c as the importance score of this channel is as follows:

[0092] I c = |γ c |

[0093] Where, γ c represents the scaling factor of the c-th channel, and I c is the importance score of this channel;

[0094] After obtaining the importance scores I1, I2,..., I c of all channels, sort them in ascending order. Set the pruning ratio p ∈ (0, 1), and mark the channels with the lowest scores as the channels to be pruned. For the channels marked as channels to be pruned, perform the following pruning operations:

[0095] In the current convolutional layer, delete the convolutional kernel parameters corresponding to the pruned channels; in the Batch Normalization layer corresponding to this convolutional layer, delete the scaling factor γ and offset β parameters of the corresponding channels; in the subsequent connected convolutional layers, delete the weight parameters in the input channels corresponding to the pruned channels; for network architectures containing residual connections or feature fusion structures, synchronously adjust the input and output channel dimensions of the corresponding paths.

[0096] S7. Perform distillation fine-tuning on the pruned student model to obtain the fine-tuned student model. The calculation formula for its distillation loss is as follows:

[0097] Ltotal = L task + αL feature + βL logits

[0098] where α and β are hyperparameters, and L task is the task loss, L feature is the feature distillation loss, and L logits is the output distillation loss. Their calculation formulas are as follows:

[0099] L task = L cls + L reg + L obj

[0100] where L cls is the classification loss, L reg is the bounding box regression loss, and L obj is the confidence loss;

[0101]

[0102] where is the feature map of the i-th layer of the student model, is the corresponding feature map of the i-th layer of the teacher model, and N is the number of selected distillation feature layers;

[0103]

[0104] Z T and Z S represent the logits outputs of the teacher model and the student model respectively. T is the temperature parameter that controls the softmax smoothness, and KL is the Kullback-Leibler divergence.

[0105] S8. Deploy the fine-tuned student model to the edge side. Based on the technical requirements of the edge side, convert the format of the model and the weight file to achieve the deployment of the model on the edge side.

[0106] The above are only the embodiments of the present invention. Specific technical solutions and / or common knowledge such as characteristics well known in the art are not described in detail herein. It should be noted that for those skilled in the art, without departing from the technical solution of the present invention, several deformations and improvements can still be made, which should also be regarded as the protection scope of the present invention, and these will not affect the implementation effect and practicality of the present invention. The protection scope required by this application should be based on the content of its claims, and the specific implementation manners described in the specification can be used to interpret the content of the claims.

Claims

1. A feature distillation tree line fault identification method based on the RT-DETR algorithm, characterized in that The recognition method steps are as follows: S1. Establish a tree line fault image dataset, including adversarial image data and labeled image data; S2. Use the adversarial image data obtained in step S1 for training, validating, and testing the SRGAN model, and use the weights with the optimal performance for super-resolution reconstruction of all images in the adversarial images to obtain super-resolution image data; S3. Divide the labeled image data obtained in step S1 and the super-resolution image data obtained in step S2 into a training set, a validation set, and a test set, and train, validate, and test the teacher model on the divided datasets respectively. After training, validating, and testing are completed, obtain the weights with the optimal performance of the teacher model; S4. Train, validate, and test the student model on the datasets divided in step S3. After training, validating, and testing are completed, obtain the weights with the optimal performance of the student model; S5. Use the teacher model obtained in step S3 to perform feature distillation on the student model to obtain a student model after feature distillation; S6. Perform pruning operations on the student model after distillation training to obtain a pruned student model; S7. Perform distillation fine-tuning on the pruned student model to obtain a fine-tuned student model; S8. Deploy the fine-tuned student model to the edge side, and based on the technical requirements of the edge side, perform format conversion on the model and weight files to achieve the deployment of the model on the edge side.

2. The feature distillation tree line fault recognition method based on the RT-DETR algorithm according to claim 1, wherein In step S1, a tree line discharge experiment platform is used to simulate the real tree line discharge phenomenon, and an image acquisition device is used to obtain image data when the tree line fault occurs, and it is divided into adversarial image data and partial labeled image data.

3. The method for identifying tree line faults by feature distillation based on the RT-DETR algorithm according to claim 1, characterized in that In step S2, in the SRGAN model, the generator G is trained using the MSE loss to generate high-resolution images, and then the discriminator is trained to learn to distinguish between real images and the generated super-resolution images. The generator is trained using adversarial loss, and the entire training process adopts an alternating training strategy until the optimal G is obtained. During training, G uses the perceptual loss function L MSE and L VGG , and its mathematical formula is: Among them, G(I LR ) i is the i-th pixel value of the super-resolution image generated by the generator G, and I HR,i is the i-th pixel value of the real high-resolution image, and N is the total number of pixels of the image; φ(·) is the intermediate layer feature map of the VGG pre-trained network, and G(I LR ) is the super-resolution image output by the generator. I HR is the real high-resolution image, and ||·|| 2 represents the L2 norm.

4. The method for identifying characteristic distillation tree line faults based on the RT-DETR algorithm according to claim 3, characterized in that, In step S2, for super-resolution reconstruction, the results of super-resolution reconstruction are evaluated through three indicators: PSNR, MSE, and SSIM, so as to obtain super-resolution image data; the calculation formulas of PSNR, MSE, and SSIM are respectively: I SR (i, j) is the pixel value of the super-resolution image, I HR (i, j) is the pixel of the true high-resolution image, and M and N are the width and height of the image respectively; Where MAX I is the maximum pixel value of the image, and MSE is the mean square error; Among them, μ SR and μ HR are the brightness of the super-resolution image and the real image, σ SR and σ HR are the contrast, σ SR,HR is the structure information, and C1 and C2 are stability factors; LPIPS(I SR ,I HR ) = ||F(I SR ) - F(I HR )|| 2 where F(·) is the feature representation extracted by the pre-trained VGG network.

5. The feature distillation tree line fault identification method based on the RT-DETR algorithm according to claim 1, characterized in that, In step S3, the RT-DETR-X model is selected as the teacher model, and in step S4, the RT-DETR-L model is selected as the student model.

6. The method for identifying tree line faults of a feature distillation based on the RT-DETR algorithm according to claim 1, characterized in that In step S6, a structured channel pruning method based on the channel scaling factor γ in the Batch Normalization layer is adopted. Specifically, for each convolutional module containing a BatchNorm layer in the student model, the scaling parameter γ corresponding to its c-th channel is extracted. c , each channel in the Batch Normalization layer contains a learnable scaling factor γ and an offset factor β, which are used to adjust the amplitude and offset of the normalized features respectively. The formula for taking the absolute value of the scaling parameter γ c as the importance score of this channel is as follows: I c = |γ c | Among them, γ c represents the scaling factor of the c-th channel, and I c is the importance score of this channel; After obtaining the importance scores I1, I2,..., I of all channels c sort them in ascending order, set the pruning ratio p ∈ (0, 1), and mark the channel with the lowest score as the channel to be pruned. For the channel marked as the channel to be pruned, perform the following pruning operation: In the current convolutional layer, delete the convolutional kernel parameters corresponding to the pruned channels; in the BatchNormalization layer corresponding to this convolutional layer, delete the scaling factor γ and offset β parameters of the corresponding channels; in the subsequent connected convolutional layers, delete the weight parameters in the input channels corresponding to the pruned channels; for network architectures containing residual connections or feature fusion structures, synchronously adjust the input and output channel dimensions of the corresponding paths.

7. The method for identifying tree line faults by feature distillation based on the RT-DETR algorithm according to claim 1, wherein In step S5, an improved global-local feature distillation is used, and its global distillation loss calculation formula is: Among them, G T is the global feature vector of the teacher model, and G S is the global feature vector of the student model, and their calculation formulas are respectively: Among them, F T (i, j) is the feature map value of the teacher model at the (i, j) position, and H and W are the height and width of the feature map respectively; where H and W are the height and width of the feature map respectively; its local distillation loss calculation formula is: L LFD = |||| L T -L S |||| 2 Among them, L T and L S represent the regional features of the teacher model and the student model respectively.

8. The method for identifying tree line faults based on the RT-DETR algorithm according to claim 1, wherein In step S7, distillation fine-tuning is adopted, and its distillation loss calculation formula is: L total = L task + αL feature + βL logits where α and β are hyperparameters, and L task is the task loss, and L feature is the feature distillation loss, and L logits is the output distillation loss, and their calculation formulas are respectively: L task = L cls + L reg + L obj Among them, L cls is the classification loss, L reg is the bounding box regression loss, and L obj is the confidence loss; Among them, is the feature map of the i-th layer of the student model, is the corresponding feature map of the i-th layer of the teacher model, and N is the number of distilled feature layers selected; Z T and Z S respectively represent the logits outputs of the teacher model and the student model, T is the temperature parameter that controls the softmax smoothness, and KL is the Kullback-Leibler divergence.

9. The feature distillation tree line fault identification method based on the RT-DETR algorithm according to claim 1, wherein In step S8, when the edge side detects a branch or a flame, it triggers an alarm and starts uploading the real-time alarm detection images to the system.

Images