A power meter image compression transmission system and transmission method based on self-encoding technology

Abstract

The present invention relates to the field of computer vision technology, and discloses a power meter image compression transmission system and transmission method based on self-encoding technology. The system includes an image acquisition and transmission device and a server. A camera is installed at the front end of the image acquisition and transmission device, and an embedded ARM chip is installed inside. A coding model is deployed on the embedded ARM chip. The output end of the image acquisition and transmission device is connected to the server through a narrowband Internet of Things, and a decoding model is deployed on the server. The transmission method is: the input power meter captures the image, which is compressed by the encoder and sent to the server via the NB‑IoT narrowband network. After the server receives the data, it decodes and restores the image. Based on this technology, the present invention solves the problems of difficult image data transmission and easy packet loss in the Internet of Things system, and reduces the power consumption of the overall monitoring system, thereby comprehensively improving the performance and improving the problems of low monitoring frequency and difficult system deployment in actual power meter monitoring.

Description

Technical Field

[0001] The present invention relates to the technical field of computer vision, and in particular to a power meter image compression transmission system and a transmission method based on a self-encoding technology. Background Art

[0002] With the development of image processing technology, power meter data analysis currently mainly uses image acquisition devices to collect and transmit images of designated meters at regular intervals and points, and then conduct manual review. Most of these devices in the early days needed to be powered by mains electricity and provide wired or WIFI networks to transmit images to servers for storage and display.

[0003] However, there are many normative requirements for environmental construction in the power field (such as the introduction of city electricity, the deployment of wired networks, etc.), and the construction is difficult. At present, the power meter monitoring device has been gradually replaced with power-free acquisition, generally using batteries or solar energy to power the equipment, and gradually using 4G networks for image transmission. There is no need for large-scale construction, and only needs to fix the image acquisition device near the designated meter. However, since the overall device is in a power-free environment, the power consumption requirements are extremely high. The camera and processor of the image acquisition device are limited to a low power range. In addition, the ordinary 4G network has high power consumption and easy packet loss when transmitting pictures. At present, the field of power meter image acquisition has begun to use dedicated low-power ARM as the core processor, which is responsible for image processing, and uses narrowband Internet of Things (NB-IOT) for image transmission, but there are the following problems: In most cases, the bandwidth and power of narrowband Internet of Things are limited, and the computing power of low-power ARM is also limited, resulting in the actual situation. The transmission of images is more difficult and the rate is slow. Often only 2 to 3 pictures can be transmitted per day, and packet loss is prone to occur, resulting in the loss of image information received by the server. To this end, we proposed a power meter image compression transmission system and transmission method based on self-encoding technology. Summary of the invention

[0004] 1. Technical issues to be solved

[0005] In view of the shortcomings of the prior art, the present invention provides an electric meter image compression transmission system and transmission method based on self-encoding technology, which has the advantages of meeting the real-time monitoring and low power consumption requirements of electric meters and achieving the transmission of up to 120 pictures per day under the same energy consumption. It solves the problems of high power consumption in picture transmission, easy packet loss during transmission, small number of pictures transmitted per day, and inability to meet the real-time monitoring requirements of electric meters.

[0006] (II) Technical solution

[0007] In order to meet the real-time monitoring and low power consumption requirements of electric meters and to achieve the purpose of transmitting more pictures every day with the same energy consumption, the present invention provides the following technical solutions: an electric meter image compression and transmission system based on self-encoding technology, comprising an image acquisition and transmission device and a server, a camera is installed at the front end of the image acquisition and transmission device, and an embedded ARM chip is installed inside, and a coding model is deployed on the embedded ARM chip, the output end of the image acquisition and transmission device is connected to the server via a narrowband Internet of Things, and a decoding model is deployed on the server.

[0008] A transmission method of the above-mentioned power meter image compression transmission system based on the self-encoding technology comprises the following steps:

[0009] a. Collect a large number of power meter images to be monitored;

[0010] b. Perform autoencoding training on the collected images, combine the characteristics of the power meter, use the gradient similarity loss function to constrain the training, and obtain two models of image compression encoding and decoding;

[0011] c. Deploy the image compression coding model to the image acquisition and transmission device based on the embedded ARM, and compress the image data collected by the front-end camera into a latent vector with a length of 256 bytes through the coding model;

[0012] d. Use narrowband IoT to transmit the 256-byte hidden vector to the server;

[0013] e. The server is deployed with a decoding model, which restores the image from the latent vector of 256 bytes to obtain the final reconstructed image for manual monitoring or subsequent data analysis.

[0014] Furthermore, in step a, the collected power meter image requires that the main target of the power meter accounts for more than 80% of the image, and all pointers and scales are clearly visible.

[0015] Furthermore, in step b, when performing the constraint training process, the gradient similarity is used to evaluate the final image reconstruction ability. First, the gradient of the image is defined as follows:

[0016]

[0017] Among them, i and j represent the position coordinates of the image data, represents the gradient component in the horizontal direction, Represents the gradient component in the vertical direction;

[0018] Then according to the above definition, the gradient is calculated as follows:

[0019] G x(i,j)=f(i+1,j)+f(i-1,j)+f(i,j+1)+f(i,j-1)-4f(i,j)

[0020] Where x(i, j) represents the image brightness data at the current position (i, j), then the gradient similarity index formula between the input image and the decoded image is:

[0021]

[0022] Furthermore, according to the defined gradient similarity, the final training loss is modified to

[0023]

[0024] in

[0025]

[0026] In the above formula, w0 and w1 are weights, with values ​​between 0 and 1. Experimental tests found that the best effect is achieved when the weight ratio is 0.3 and 0.7. p is the distance level, with a value of 2.

[0027] Furthermore, according to the final training loss function, the difference between the decoded image and the input image is evaluated, and the encoding and decoding model of the power meter is trained using the back propagation algorithm.

[0028] Furthermore, based on the obtained power meter encoding and decoding model, the encoding model is deployed to the embedded ARM system. The system uses the encoding model to compress and encode the collected camera data at fixed intervals to obtain a 256-byte latent vector, and then sends the vector data to the server through the serial port network module. The server deploys the obtained decoding model, and after receiving the vector data, it decodes the image to generate the final image and stores it.

[0029] Furthermore, in step b, the collected image is trained by self-encoding using a convolutional neural network. The network structure is an encoder network. All parameters in the encoder network use a power series of 2, remove the maximum pooling in the traditional convolution method, and directly use convolution to achieve feature downsampling function; the decoder network is cascaded with the convolution layer parameters of the corresponding encoder in the deconvolution layer, and all convolution operations use a depth-separable convolution calculation method to reduce the amount of calculation.

[0030] Furthermore, the calculation method of the depthwise separable convolution is as follows: 8-bit data is used for model quantization at the encoder end. The quantization method is to use a predefined 1000 training set images. After the encoding model is calculated, the maximum value that appears during the forward operation of each layer of the convolutional neural network is counted, and then the maximum value of each layer is normalized with the model parameters, and represented by 8 bits to obtain the final quantized encoding model.

[0031] (III) Beneficial effects

[0032] Compared with the prior art, the present invention provides a power meter image compression transmission system and transmission method based on self-encoding technology, which has the following beneficial effects:

[0033] In the present invention, based on the self-encoding framework, the image data collected by the camera is compressed and encoded, and only the NB-IoT network is needed to transmit high-quality images. At the same time, an image decoding algorithm is deployed on the server side, and only a few hundred bytes are needed to transmit the original VGA image with a resolution of 640x480. The restored SIMM can reach 86.01. Compared with the traditional method, the present invention solves the problems of slow transmission rate, high packet loss rate, long time consumption, high energy consumption and the like of image data transmission under narrowband Internet of Things, and meets the real-time monitoring and low power consumption requirements of power meters. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 The overall system flow of a power meter image compression transmission system and a transmission method based on self-encoding technology proposed by the present invention;

[0035] Figure 2 This is a block diagram of the working principle of a power meter image compression transmission system based on self-encoding technology proposed by the present invention;

[0036] Figure 3 This is a schematic diagram of the self-encoding network architecture in a power meter image compression transmission system and transmission method based on the self-encoding technology proposed by the present invention;

[0037] Figure 4 The invention proposes an electric meter image compression transmission system and transmission method based on auto-encoding technology and an electric meter image auto-encoding training loss curve. DETAILED DESCRIPTION

[0038] The technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. 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.

[0039] Reference Figure 1-4 , an electric power meter image compression transmission system based on self-encoding technology, comprising an image acquisition and transmission device and a server, wherein a camera is installed at the front end of the image acquisition and transmission device, and an embedded ARM chip is installed inside, and a coding model is deployed on the embedded ARM chip, and an output end of the image acquisition and transmission device is connected to the server through a narrowband Internet of Things, and a decoding model is deployed on the server;

[0040] The above system specifically uses Broadcom BCM2835 low-power ARM chip as the core chip. The main frequency of BCM2835 is 1GHz, the RAM is 512MB, the CSI interface is used to access the ov5640 camera, and the image acquisition resolution is set to 640x480; CN18S is used as the NB-Iot communication chip, and the core board transmits data with the communication module through the serial port; STC8A is used as the main control chip for system power management. Figure 2 The block diagram of the working principle of the system is shown in Figure 2.

[0041] The transmission method of the power meter image compression transmission system based on the self-encoding technology comprises the following steps:

[0042] a. Collect a large number of images of power meters to be monitored, requiring that the main target of the power meter accounts for more than 80% of the image, and all pointers and scales are clearly visible;

[0043] b. Perform auto-encoding training on the collected images. See the schematic diagram of the auto-encoding network architecture for details. Figure 3 , the collected images are trained by convolutional neural network for autoencoding, and the network structure is encoder network, see the table below:

[0044]

[0045]

[0046] The linear combination of L2 distance and gradient similarity index is used to evaluate the network reconstruction ability. The final loss function defined in this example is:

[0047]

[0048] in,

[0049] G x (i,j)=x(i+1,j)+x(i-1,j)+x(i,j+1)+x(i,j-1)-4x(i,j)

[0050]

[0051] They are defined as the gradient similarity index of the input image and the reconstructed image, x and It is the brightness data of the input image and the reconstructed image.

[0052] The optimizer for network training is Adam optimization, and the initial learning rate is defined as 0.0001. The typical training loss curve is shown in Figure 4 .

[0053] Generally speaking, when the loss drops below 0.05, the encoding and decoding model is considered usable.

[0054] The coding model was quantized, and the predefined 1,000 training set images were used. After the coding model was calculated, the maximum value of each layer of the convolutional neural network forward operation was counted, and then the maximum value of each layer and the model parameters were normalized and represented by 8 bits to obtain the final quantized coding model. Compared with the unquantized model, the operation efficiency was improved by 6.5 times, and the structural similarity index of all the images in the verification set after reconstruction remained above 86%, meeting the actual application needs.

[0055] The quantized encoding model is then deployed to the BCM2835 platform, where the convolution calculation uses deep separable convolution, and the multiplication and addition operation is accelerated using ARM's Neon instruction. The final encoding time for a single image is 3 to 6 seconds. Finally, the encoded 256-byte data is transmitted to the NB-IoT chip CN18S through the serial port, and then sent to the server through the narrowband Internet of Things.

[0056] The server deploys the trained decoding model and performs reverse decoding after receiving the 256-byte encoded data. Decode, reconstruct the acquired images and store them for subsequent task processing. The overall framework of the system is shown in Figure 1 , where q(z|x) is the feature distribution model under a given input sample x, i.e., the encoding model; q(z|x) is the prediction distribution model under a given input sample x, i.e., the decoding model; q(z) is the feature vector distribution, i.e., the output after encoding, the input of the decoding model.

[0057] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A power meter image compression transmission system based on self-encoding technology, including an image acquisition and transmission device and a server, characterized in that The front end of the image acquisition and transmission device is equipped with a camera, and an embedded ARM chip is installed inside, and a coding model is deployed on the embedded ARM chip. The output end of the image acquisition and transmission device is connected to the server through a narrowband Internet of Things, and a decoding model is deployed on the server. The specific implementation steps are as follows: a. Collect a large number of power meter images to be monitored; b. Perform autoencoding training on the collected images, combine the characteristics of the power meter, use the gradient similarity loss function to constrain the training, and obtain two models of image compression encoding and decoding; c. Deploy the image compression coding model to the image acquisition and transmission device based on the embedded ARM, and compress the image data collected by the front-end camera into a latent vector with a length of 256 bytes through the coding model; d. Use narrowband IoT to transmit the 256-byte hidden vector to the server; e. The server is deployed with a decoding model, which restores the image from the latent vector of 256 bytes to obtain the final reconstructed image for manual monitoring or subsequent data analysis; In step a, the collected power meter image requires that the main target of the power meter accounts for more than 80% of the image, and all pointers and scales are clearly visible; In step b, during the constraint training process, gradient similarity is used to evaluate the final image reconstruction capability. First, the gradient of the image is defined as follows: ； Among them, i and j represent the position coordinates of the image data, represents the gradient component in the horizontal direction, Represents the gradient component in the vertical direction; Then according to the above definition, the gradient is calculated as follows: ； Where X(i, j) represents the image brightness data at the current position (i, j), then the gradient similarity index formula between the input image and the decoded image is: ; According to the defined gradient similarity, the final training loss is modified to ； in ； In the above formula, and is the weight, which is between 0 and 1. Experimental tests show that the best weight ratio is 0.3 and 0.

7. p is the distance level, which is 2. According to the final training loss function, the difference between the decoded image and the input image is evaluated, and the encoding and decoding model of the power meter is trained using the back propagation algorithm.

2. According to claim 1, a power meter image compression transmission system based on self-encoding technology is characterized in that ,According to the obtained power meter encoding and decoding model, the encoding model is deployed to the embedded ARM system.,The system uses the encoding model to compress and encode the collected images at a fixed time interval,,obtaining a 256-byte latent vector, and then sends the vector data to the server through the serial port network module.,The server deploys the obtained decoding model, and after receiving the vector data,,it will decode the image, generate the final image and store it.

3. The power meter image compression transmission system based on the self-encoding technology according to claim 2 is characterized in that In step b, the collected image is trained by convolutional neural network for self-encoding. The network structure is a codec network, in which the encoding end and the decoding end are symmetrically distributed, but the decoding end also cascades the feature information extracted by the encoding end to enhance the image restoration ability. All parameters in the codec network use the power series of 2, remove the maximum pooling in the traditional convolution method, and directly use convolution to realize the feature downsampling function.

4. The power meter image compression transmission system based on the self-encoding technology according to claim 3 is characterized in that: The decoder network in the codec network is cascaded with the convolution layer parameters of the corresponding encoder in the deconvolution layer, and all convolution operations use a depth-separable convolution calculation method, specifically: 8-bit data is used for model quantization at the encoder end, and the quantization method is to use a predefined 1000 training set images, and after the encoding model is calculated, the maximum value of each layer of the convolutional neural network during the forward operation is counted, and then the maximum value of each layer and the model parameters are normalized, and 8 bits are used to represent it to obtain the final quantized coding model.

Images