Intelligent data transmission method based on comparison self-supervision

Abstract

The invention belongs to the technical field of data processing, and particularly relates to an intelligent data transmission method based on comparison self-supervision, which comprises the following steps: sorting and integrating parking picture data and video data in different scenes to obtain a parking image data set, and obtaining a positive sample and a negative sample by adopting a strong and weak data enhancement method; the method comprises the steps that firstly, a comparison self-supervision module, a rotation self-supervision module and an image reconstruction module are trained, then a double-encoder-decoder frame is constructed, then the double-encoder-decoder frame is optimized, the comparison self-supervision module, the rotation self-supervision module and the image reconstruction module are trained respectively, and the optimized double-encoder-decoder frame is obtained; then parking data are input into the optimized double encoder-decoder framework, and the integrity and reliability of transmission data are ensured while the compression ratio is greatly improved; meanwhile, the compression mode can be dynamically selected according to the network state, and intelligent data transmission is achieved.

Description

Technical Field

[0001] The invention belongs to the technical field of data processing, and in particular relates to an intelligent data transmission method based on comparative self-supervision. Background Art

[0002] Intelligent transportation is the application of artificial intelligence technology in the field of transportation and autonomous driving. As a key component of intelligent transportation, efficient and reliable data transmission is essential. Modern parking algorithm cameras can capture high-definition images and even video evidence, which is essential for accurate parking management and potential legal evidence collection.

[0003] The current methods for parking data transmission cannot effectively cope with complex and changeable network scenarios and unstable network environments. On the one hand, high-definition image and video data are usually large in size and require a lot of bandwidth to achieve fast and stable transmission. However, in actual application scenarios, network conditions are often unsatisfactory. Network coverage in urban environments may have blind spots or face congestion during peak hours, resulting in insufficient available bandwidth. The large amount of data transmission under such bandwidth limitations seriously affects the transmission efficiency and may cause delays or loss of key information. On the other hand, in complex urban environments, network connections may fluctuate or be interrupted due to various factors. This instability not only affects the continuity of data transmission, but may also cause data packet loss or errors, thereby affecting the integrity and reliability of parking evidence. Traditional data compression methods, such as JPEG compression for images and H.264 compression for videos, can reduce the amount of data to a certain extent, but it is still challenging to achieve high compression rates while maintaining high quality. Moreover, these methods are usually static and cannot dynamically adjust the compression strategy according to real-time network conditions, making it difficult to adapt to changes in network conditions. Summary of the invention

[0004] In order to overcome the shortcomings of the above-mentioned prior art, the present invention provides an intelligent data transmission method based on comparative self-supervision, constructs a dual encoder-decoder framework based on comparative self-supervision, intelligently selects the appropriate compression mode according to the network conditions, realizes efficient compression and intelligent transmission of data, thereby ensuring the efficiency and reliability of parking data transmission.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] An intelligent data transmission method based on contrastive self-supervision comprises the following steps: S1. Data acquisition: The parking image data and video data in different scenes are sorted and integrated to obtain the parking image dataset, and the strong and weak data enhancement methods are used to obtain positive samples and negative samples; S2. Dual encoder-decoder framework construction: construct a dual encoder-decoder framework, the dual encoder-decoder framework includes a first encoder, a second encoder, a first decoder, a second decoder, a contrast self-supervision module, a rotation self-supervision module and an image reconstruction module; S3, dual encoder-decoder framework optimization: train the contrast self-supervision module, rotation self-supervision module and image reconstruction module respectively, and obtain the optimized dual encoder-decoder framework through the training of the three modules; S4, data transmission: input the parking data into the optimized dual encoder-decoder framework to realize intelligent data transmission.

[0007] As a further technical solution of the present invention, the process of obtaining positive samples and negative samples in step S1 is as follows: for a sample x in the input sample set X, after strong and weak data enhancement, and Get two positive sample sets respectively and , negative sample set From other input samples Q: , , .

[0008] As a further technical solution of the present invention, the working process of the comparison self-supervision module described in step S2 is: inputting the positive and negative samples into the first encoder and the second encoder to extract high-level semantic features, obtaining the feature vector of the samples, and performing similarity measurement between the positive samples and the negative samples; the working process of the rotation self-supervision module is: first randomly rotating the original parking image by 0°, 90°, 180° or 270°, and the rotated original parking image is sequentially passed through the first encoder and the second encoder to obtain feature representation, and then the rotation prediction classification task is performed, and the rotation angle of the application is predicted by the 4 classifier heads; the working process of the image reconstruction module is: first inputting the image sample into the first encoder to generate a high-dimensional feature vector, the high-dimensional feature vector enters the second encoder to be compressed into a low-dimensional feature vector, the second decoder decodes the low-dimensional feature vector into an intermediate feature representation, and the first decoder reconstructs the intermediate feature representation into a complete image.

[0009] As a further technical solution of the present invention, the specific process of step S3 training the contrastive self-supervision module is: inputting positive and negative samples of sample x into the encoder, and performing contrastive self-supervision training after measuring the similarity. During training, the contrastive learning loss function is calculated as follows: , in, Represents the metric function used to calculate the similarity between the feature vectors of sample images. The metric function uses inner product or cosine similarity. The parameter α is the weight coefficient, which is used to adjust the relative importance of strong and weak enhanced positive samples in the training process. The function Two implementation methods are used to cope with different training requirements and network bandwidth conditions: At high bandwidth, Represented as passing only the first encoder The extracted high-dimensional feature vector, i.e. ; In low bandwidth environments, First pass through the first encoder Then through the second encoder The extracted low-dimensional feature vector is .

[0010] As a further technical solution of the present invention, the specific process of step S3 training the rotation self-supervision module is: Input sample x, first randomly rotate the original parking image x by 0°, 90°, 180° or 270° to obtain the rotated image , and then rotate the image Input first encoder , get high-dimensional features , and then the high-dimensional features Enter the second encoder , get low-dimensional features ; Then the high-dimensional features and low-dimensional features Perform residual connection and input a projection head composed of a multi-layer perceptron. The output of the projection head After the classifier is used for 4 classifications, the predicted probability of each angle is obtained. , the rotation self-supervision module is trained to learn the rotation-invariant features and the geometric characteristics of the image rotation. The prediction loss during training is calculated as: , by summing the 4 angle categories, the cross entropy loss is calculated to calculate the difference between the predicted angle and the true angle; It is the one-hot encoding of the true label, where the position corresponding to the true angle is 1 and the other positions are 0.

[0011] As a further technical solution of the present invention, the specific process of step S3 training the image reconstruction module is as follows: For the input sample x, first input the first encoder , get high-dimensional features ; then Input encoder , to obtain more compact low-dimensional features , then Input two additional independent multi-layer perceptrons , get the mean of the feature space and variance . Then use the reparameterization technique to sample from Sampling latent variables from the distribution ; The latent variable Input decoder , and reconstruct high-dimensional features ; Finally, the high-dimensional features Input decoder , and obtain the reconstructed image , three loss functions are used in the training process: , , ,in and Represent the reconstruction losses of feature reconstruction and image reconstruction at a fine-grained level, is the KL divergence loss, the total loss of image reconstruction ,in , and is the weight coefficient, so the total loss function of the dual encoder-decoder framework is as follows: .

[0012] As a further technical solution of the present invention, a bandwidth adaptive selector is provided between the encoder and the decoder, and the bandwidth adaptive selector numbers the transmitted data blocks, and selects different data compression methods and transmits them according to different bandwidth states, specifically: when the network bandwidth is in good condition, the first encoder is used to compress the original image data to obtain a high-dimensional feature vector, and the first decoder reconstructs the high-dimensional image according to the high-dimensional feature vector; if the network bandwidth is limited, the first encoder and the second encoder are used locally at the same time to obtain a low-dimensional feature vector, and the low-dimensional feature vector enters two independent multi-layer perceptrons to calculate the mean and variance of the feature space, and then maps out new latent variables, and the latent variables are reconstructed by the second decoder and the first decoder; if the network bandwidth is completely unable to complete the communication, the transmitted data is temporarily stored locally, and the data transmission continues after the network communication is restored.

[0013] Compared with the prior art, the present invention has the following beneficial effects: The present invention is based on a dual encoder-decoder framework with comparative self-supervision, which can dynamically select a compression method according to network bandwidth conditions, thereby greatly improving the compression rate while ensuring the complete reliability of transmitted data; at the same time, it can dynamically select a compression method according to network status to achieve intelligent data transmission. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0015] Figure 1 A detailed flow chart of an intelligent data transmission method based on comparative self-supervision in an embodiment of the present invention is provided.

[0016] Figure 2 It is a framework flow chart of the intelligent data transmission method based on comparative self-supervision in an embodiment of the present invention.

[0017] Figure 3 Optimization flow chart of the dual encoder-decoder framework in an embodiment of the present invention.

[0018] Figure 4 Schematic diagram of the encoder structure in an embodiment of the present invention.

[0019] Figure 5 Schematic diagram of the decoder structure in an embodiment of the present invention. DETAILED DESCRIPTION

[0020] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0021] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0022] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0023] Introduction to contrastive self-supervised learning: Compared with supervised learning and semi-supervised learning in deep learning, contrastive self-supervised learning is an innovative method of unsupervised learning. Its uniqueness lies in the fact that it does not require any prior knowledge or labeled data, but can learn directly from raw data. This method cleverly uses the similarity measure between positive and negative samples and combines other self-supervised techniques (such as rotation prediction tasks and reconstruction of occluded areas) to deeply explore the rich features and high-level semantic information inherent in data samples.

[0024] The core advantages of self-supervised learning are its ability to: automatically model sample features without manual labeling; capture complex patterns and potential structures in the data; and learn more universal and robust feature representations.

[0025] This learning method can not only effectively cope with traditional classification tasks, but also demonstrates excellent performance in image recognition, natural language processing, and cross-modal learning. By making full use of a large amount of unlabeled data, contrastive self-supervised learning provides an efficient solution to practical problems such as data scarcity and high annotation costs.

[0026] Embodiment 1: like Figure 1-5 As shown, this embodiment provides an intelligent data transmission method based on contrastive self-supervision, which uses the original sample data to train an end-to-end model that can effectively compress image data and restore and reconstruct images. Based on contrastive self-supervision learning, three auxiliary task modules, namely, contrastive self-supervision module, rotation self-supervision module and image reconstruction module, are designed, so that the proposed dual encoder-decoder framework can obtain feature representation ability and information modeling ability from the data, and then perform dimensionality reduction compression and restoration and reconstruction on the input image to cope with unstable, complex and changeable network bandwidth conditions. Specifically, the following steps are included: S1. Data acquisition: The parking image data and video data in different scenes are sorted and integrated to obtain the parking image dataset. The strong and weak data enhancement methods are used to obtain positive and negative samples. For a sample x in the input sample set X, after strong and weak data enhancement, and Get two positive sample sets respectively and , negative sample set From other input samples Q: , , .

[0027] S2. Dual encoder-decoder framework construction: construct a dual encoder-decoder framework, the dual encoder-decoder framework includes a first encoder , Second Encoder , the first decoder , the second decoder , contrast self-supervision module, rotation self-supervision module and image reconstruction module; wherein the working process of the contrast self-supervision module is: inputting positive and negative samples into the first encoder and the second encoder Perform high-level semantic feature extraction to obtain the feature vector of the sample and measure the similarity between the positive sample and the negative sample. The working process of the rotation self-supervision module is as follows: first, the original parking image is randomly rotated by 0°, 90°, 180° or 270°, and the rotated original parking image is sequentially passed through the first encoder and the second encoder Get the feature representation, then perform the rotation prediction classification task, and predict the rotation angle of the application through the 4 classifier heads; the working process of the image reconstruction module is: first input the image sample into the first encoder Generate high-dimensional feature vectors, high-dimensional feature vectors Enter the second encoder Compressed into a low-dimensional feature vector , the second decoder The low-dimensional feature vector Decoded into the intermediate feature representation z, the first decoder Reconstruct the intermediate feature representation z into a complete image; A bandwidth adaptive selector is provided between the encoder and the decoder. The bandwidth adaptive selector numbers the transmitted data blocks and selects different data compression methods for transmission according to different bandwidth conditions. Specifically, when the network bandwidth is good, the first encoder is used. The original image data is compressed to obtain a high-dimensional feature vector. The first decoder High-quality image reconstruction is performed based on the high-dimensional feature vector; if the network bandwidth is limited, the first encoder is used locally at the same time and the second encoder The low-dimensional feature vector is obtained, and the low-dimensional feature vector enters two independent multi-layer perceptrons to calculate the mean and variance of the feature space, and then maps out new latent variables. The latent variables pass through the second decoder and the first decoder Image reconstruction is performed; if the network bandwidth is completely unable to complete the communication, the transmitted data is temporarily stored locally and the data transmission continues after the network communication is restored.

[0028] S3, dual encoder-decoder framework optimization: train the contrastive self-supervision module, rotation self-supervision module and image reconstruction module respectively. Through the training of the three modules, the optimized dual encoder-decoder framework is obtained. The specific process of training the contrastive self-supervision module is as follows: input the positive and negative samples of sample x into the encoder, measure the similarity and then perform contrastive self-supervision training. During training, the contrastive learning loss function is calculated as follows: , in, It represents the metric function used to calculate the similarity between the feature vectors of sample images. The metric function uses inner product or cosine similarity. The parameter α is the weight coefficient, which is used to adjust the relative importance of strong and weak enhanced positive samples in the training process. The function Two implementation methods are used to cope with different training requirements and network bandwidth conditions: At high bandwidth, Represented as passing only the first encoder The extracted high-dimensional feature vector, i.e. ; In low bandwidth environments, First pass through the first encoder Then through the second encoder The extracted low-dimensional feature vector is ; According to the contrastive learning method of different dimensions, two contrastive learning loss functions are calculated simultaneously to optimize the framework. The contrastive learning loss function is designed to guide the model to learn a high-quality feature representation space, in which the model is motivated to map semantically similar samples (original samples and their enhanced versions) to adjacent positions, while mapping semantically irrelevant samples (negative samples) to distant positions, effectively promoting the model to capture the essential semantic features of the data, thereby constructing a feature representation that is both discriminative and expressive. This method not only improves the model's ability to understand the intrinsic structure of the data, but also enhances its generalization performance in various downstream tasks, especially in scenarios where labeled data is scarce. Through this self-supervision method, the model can extract valuable information from a large amount of unlabeled data, laying a solid foundation for subsequent specific tasks.

[0029] The specific process of training the rotation self-supervision module is:

[0030] Input sample x, first randomly rotate the original parking image x by 0°, 90°, 180° or 270° to obtain the rotated image , and then rotate the image Input first encoder , get high-dimensional features , and then the high-dimensional features Enter the second encoder , get low-dimensional features ; Then the high-dimensional features and low-dimensional features Perform residual connection and input a projection head composed of a multi-layer perceptron. The output of the projection head After the classifier is used for 4 classifications, the predicted probability of each angle is obtained. , the rotation self-supervision module is trained to learn the rotation-invariant features and the geometric characteristics of the image rotation. The prediction loss during training is calculated as: , by summing the 4 angle categories, the cross entropy loss is calculated to calculate the difference between the predicted angle and the true angle; It is the one-hot encoding of the true label, where the position corresponding to the true angle is 1 and the other positions are 0.

[0031] The specific process of training the image reconstruction module is: For the input sample x, first input the first encoder , get high-dimensional features ; then Input encoder , to obtain more compact low-dimensional features , then Input two additional independent multi-layer perceptrons , get the mean of the feature space and variance . Then use the reparameterization technique to sample from Sampling latent variables from the distribution ; The latent variable Input decoder , and reconstruct high-dimensional features ; Finally, the high-dimensional features Input decoder , and obtain the reconstructed image , three loss functions are used in the training process: , , ,in and Represent the reconstruction losses of feature reconstruction and image reconstruction at a fine-grained level, respectively. Through feature and image reconstruction losses, it is possible to learn to accurately capture and reproduce the visual content of the input image; It is the KL divergence loss. The KL divergence loss encourages the model to learn a smooth and continuous latent space, which is conducive to image interpolation and semantic operations. The total loss of image reconstruction is ,in , and is the weight coefficient, so the total loss function of the dual encoder-decoder framework is as follows: ,Through this diversified training strategy, the framework achieves all-round feature learning, ,which enables the framework to achieve intelligent data compression and ,efficient transmission when processing parking data.

[0032] S4, data transmission: input the parking data into the optimized dual encoder-decoder framework to realize intelligent data transmission.

[0033] In this embodiment, when the network bandwidth is good, the first encoder is used. The original image data is compressed. The original image data is 1920 × 1080 × 3, and the high-dimensional feature is 512 × 512. Each feature point is saved in float32 format, that is, the size of a feature point is 4 bytes, and the data volume is reduced by about 83.1%. When the network bandwidth is limited, the first encoder is used locally at the same time. and the second encoder The original image data is 1920 × 1080 × 3, the high-dimensional feature is 256 × 256, and each feature point is saved in float32 format, that is, the size of a feature point is 4 bytes, and the data volume is reduced by about 95.8%.

[0034] The structures and algorithms not described in detail in this article are general techniques in the field.

[0035] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

[0036] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. An intelligent data transmission method based on contrastive self-supervision, characterized in that: The following steps are involved: S1. Data acquisition: The parking image data and video data in different scenes are sorted and integrated to obtain the parking image dataset, and the strong and weak data enhancement methods are used to obtain positive samples and negative samples; S2. Dual encoder-decoder framework construction: construct a dual encoder-decoder framework, the dual encoder-decoder framework includes a first encoder, a second encoder, a first decoder, a second decoder, a contrast self-supervision module, a rotation self-supervision module and an image reconstruction module; S3, dual encoder-decoder framework optimization: train the contrast self-supervision module, rotation self-supervision module and image reconstruction module respectively, and obtain the optimized dual encoder-decoder framework through the training of the three modules; S4, data transmission: input the parking data into the optimized dual encoder-decoder framework to realize intelligent data transmission.

2. According to claim 1, the intelligent data transmission method based on contrastive self-supervision is characterized in that: Step S1 The process of obtaining positive and negative samples is as follows: for a sample x in the input sample set X, after strong and weak data enhancement and Get two positive sample sets respectively and , negative sample set From other input samples Q: , , .

3. According to the intelligent data transmission method based on contrastive self-supervision as described in claim 2, the working process of the contrastive self-supervision module in step S2 is: inputting positive and negative samples into the first encoder and the second encoder to extract high-level semantic features, obtaining feature vectors of the samples, and performing similarity measurement between the positive samples and the negative samples; the working process of the rotation self-supervision module is: first randomly rotating the original parking image by 0°, 90°, 180° or 270°, the rotated original parking image is sequentially passed through the first encoder and the second encoder to obtain feature representation, and then the rotation prediction classification task is performed, and the rotation angle of the application is predicted by the 4 classifier heads; the working process of the image reconstruction module is: first inputting the image sample into the first encoder to generate a high-dimensional feature vector, the high-dimensional feature vector enters the second encoder to be compressed into a low-dimensional feature vector, the second decoder decodes the low-dimensional feature vector into an intermediate feature representation, and the first decoder reconstructs the intermediate feature representation into a complete image.

4. According to the intelligent data transmission method based on contrastive self-supervision according to claim 3, the specific process of step S3 training the contrastive self-supervision module is: inputting positive and negative samples of sample x into the encoder, and performing contrastive self-supervision training after measuring the similarity. During training, the contrastive learning loss function is calculated as follows: ， in, Represents the metric function used to calculate the similarity between the feature vectors of sample images. The metric function uses inner product or cosine similarity. The parameter α is the weight coefficient, which is used to adjust the relative importance of strong and weak enhanced positive samples in the training process. The function Two implementation methods are used to cope with different training requirements and network bandwidth conditions: At high bandwidth, Represented as passing only the first encoder The extracted high-dimensional feature vector, i.e. ; In low bandwidth environments, First pass through the first encoder Then through the second encoder The extracted low-dimensional feature vector is .

5. According to the intelligent data transmission method based on contrast self-supervision according to claim 4, the specific process of step S3 training the rotation self-supervision module is: Input sample x, first randomly rotate the original parking image x by 0°, 90°, 180° or 270° to obtain the rotated image , and then rotate the image Input first encoder , get high-dimensional features , and then the high-dimensional features Enter the second encoder , get low-dimensional features ; Then the high-dimensional features and low-dimensional features Perform residual connection and input a projection head composed of a multi-layer perceptron. The output of the projection head After the classifier is used for 4 classifications, the predicted probability of each angle is obtained. , the rotation self-supervision module is trained to learn the rotation-invariant features and the geometric characteristics of the image rotation. The prediction loss during training is calculated as: , by summing the 4 angle categories, the cross entropy loss is calculated to calculate the difference between the predicted angle and the true angle; It is the one-hot encoding of the true label, where the position corresponding to the true angle is 1 and the other positions are 0.

6. According to the intelligent data transmission method based on contrastive self-supervision in claim 5, the specific process of step S3 training the image reconstruction module is: For the input sample x, first input the first encoder , get high-dimensional features ; then Input encoder , to obtain more compact low-dimensional features , then Input two additional independent multilayer perceptrons , get the mean of the feature space and variance . Then use the reparameterization technique to sample from Sampling latent variables from the distribution ; The latent variable Input Decoder , and reconstruct high-dimensional features ; Finally, the high-dimensional features Input Decoder , and obtain the reconstructed image , three loss functions are used in the training process: , , ,in and Represent the reconstruction losses of feature reconstruction and image reconstruction at a fine-grained level, is the KL divergence loss, the total loss of image reconstruction ,in , and is the weight coefficient, so the total loss function of the dual encoder-decoder framework is as follows: .

7. According to the intelligent data transmission method based on contrastive self-supervision in claim 6, a bandwidth adaptive selector is provided between the encoder and the decoder, and the bandwidth adaptive selector numbers the transmitted data blocks, and selects different data compression methods and transmits them according to different bandwidth states, specifically: when the network bandwidth is in good condition, the first encoder is used to compress the original image data to obtain a high-dimensional feature vector, and the first decoder reconstructs the high-dimensional image according to the high-dimensional feature vector; if the network bandwidth is limited, the first encoder and the second encoder are used locally at the same time to obtain a low-dimensional feature vector, and the low-dimensional feature vector enters two independent multi-layer perceptrons to calculate the mean and variance of the feature space, and then maps out new latent variables, and the latent variables are reconstructed by the second decoder and the first decoder; if the network bandwidth is completely unable to complete the communication, the transmitted data is temporarily stored locally, and the data transmission continues after the network communication is restored.

Images