Visual laser data feature compression and extraction method and system

Abstract

The invention discloses a visual laser data feature compression and extraction method and system, and relates to the technical field of feature compression and extraction, and the method comprises the following steps: building a feature compression frame based on a convolutional neural network, and extracting an intermediate feature; according to statistical features of the features, determining redundant non-informational features or non-informational features with less information amount, and deleting features most similar to the non-informational features; discrete cosine transform is introduced to further compress statistical redundancy, and redundancy between feature channels is eliminated through feature rearrangement, feature transformation and feature quantization; in the receiver, the removed features are recovered from the features of the subset of non-informative features. According to the method, the features in the similarity subset are deleted in the encoder and are recovered according to the non-information subset in the decoder, so that the pressure of data transmission and storage is remarkably reduced, and meanwhile, the integrity and effectiveness of feature information are ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of feature compression and extraction, and in particular to a method and system for visual laser data feature compression and extraction. Background Art

[0002] In the application of digital twin technology for power transmission and transformation equipment, smart devices equipped with visual / laser sensors have been widely used to provide various image videos or laser data, which will be transmitted and stored to remote servers to perform various analysis and monitoring tasks, such as image classification and defect detection. How to economically and efficiently transmit and store massive amounts of data is a huge challenge. Traditional multimodal data compression methods aim to maintain good human-perceivable quality. However, with the development of cloud services and data sharing platforms, real-time status detection of power equipment has also put forward higher requirements for the transmission and storage of multimodal data of power transmission and transformation equipment.

[0003] Deep neural networks can learn feature representations by training on large-scale data without the need for hand-crafted feature extractors. Therefore, a paradigm of encoding and transmitting intermediate features can be adopted. The neural network is divided into two parts, deployed on the edge side and the cloud side respectively. The front end of the network extracts intermediate features on the edge side, and the rest of the network checks the task-specific accuracy through the features received in the cloud. In order to obtain higher feature compression performance, the redundancy between feature channels can be eliminated through feature rearrangement, feature transformation, and feature quantization. Summary of the invention

[0004] In view of the problems existing in the existing visual laser data feature compression and extraction and system, the present invention is proposed.

[0005] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0006] In a first aspect, an embodiment of the present invention provides a method for visual laser data feature compression and extraction, which comprises the following steps:

[0007] Establish a feature compression framework based on convolutional neural networks to extract intermediate features;

[0008] According to the statistical characteristics of the features, redundant or low-information uninformative features are determined, and the features that are most similar to the uninformative features are deleted;

[0009] Discrete cosine transform is introduced to further compress statistical redundancy, and redundancy between feature channels is eliminated through feature rearrangement, feature transformation and feature quantization;

[0010] In the receiver, the removed features are restored based on the features of the non-informative feature subset.

[0011] As a preferred solution of the visual laser data feature compression and extraction method of the present invention, the specific steps also include:

[0012] The input image is set to x, and the features obtained from the N middle layers of the network form a feature tensor space whose size is set to H×C×W, where C, H, and W represent the number of channels, height, and width of the feature, respectively. Before the nonlinear activation function, the output feature f(l) of the lth layer is given by the formula, which is expressed as:

[0013] f(l)=N(x,l)=∑h(l,i)*f(l-1,i)+α(l,i);

[0014] Where h(l,i) represents the convolution parameter of the i-th channel of the l-th layer, α(l,i) is the bias parameter of the convolutional neural network, and * represents the 2D convolution operation;

[0015] Determine the features with the least relevance to the monitoring task to form a non-informative subset S n , according to the similarity metric, select the features that are most similar to the non-informative features to form a similarity subset S s , S s and S n The features in have a one-to-one correspondence, and then all the features are divided into three subsets;

[0016] S=S n ∪S i ∪S s ={f(l,i),i∈[1,C]};

[0017] In the formula, S represents the entire feature, S i is the set of remaining features. S s The features in will be removed in the encoder;

[0018] Output feature S ψ It is expressed as:

[0019] S ψ =Ψ(S);

[0020] In the receiver, the S n The corresponding features in restore the removed features;

[0021] S ψ The transformation is performed by transforming Γ[·]. Finally, the discrete code y(l) is generated by quantization and entropy coding, which is represented by C[·]:

[0022] y(l)=C(Γ(S ψ ),θ C );

[0023] In the formula, θC Indicates the encoder parameter set to be tuned;

[0024] In the decoder, reconstruct the features:

[0025]

[0026] In the formula, It is the inverse process of the corresponding process in the encoder.

[0027] As a preferred solution of the visual laser data feature compression and extraction method of the present invention, the steps of extracting intermediate features and determining phase velocity include:

[0028] The intermediate features include: non-information features, target information features and similarity feature subsets;

[0029] The subset of non-informative features is denoted as: S n ={f(l,i dx )|i dx ∈1,K};

[0030] Among them, i dx is the index of the non-informative feature channel;

[0031] Therefore, the similarity of non-informative features is expressed as:

[0032]

[0033] Among them, ρ i,j represents the similarity between f(l,i) and f(l,j), ρ i,j Determined by Pearson correlation coefficient, mutual information or k2 norm;

[0034] ρ i,j =||f(i,j)-f(l,j)||2.

[0035] As a preferred solution of the visual laser data feature compression and extraction method of the present invention, wherein: the discrete cosine transform is expressed as,

[0036] z f (l) = A T f′(l)A;

[0037] In the formula, z f (l) is the transformed feature data, A is the transformation matrix, [·] T is the matrix transpose.

[0038] As a preferred solution of the visual laser data feature compression and extraction method of the present invention, the step of eliminating redundancy between feature channels includes:

[0039] Without loss of generality, uniform quantization is used and n-bit quantization is performed as follows:

[0040]

[0041] In the formula, max(z f (l)) and min(z f (l)) are the maximum and minimum values ​​in the feature data, [z q (l)] H,W is the quantized feature data, Indicates rounding down to the nearest integer. q (l) will be input into the arithmetic coding engine to further eliminate its statistical redundancy.

[0042] As a preferred embodiment of the visual laser data feature compression and extraction method of the present invention, the step of restoring representative features includes:

[0043] The feature extraction decoder includes: entropy decoding, inverse quantization, inverse transformation and feature recovery;

[0044] Among them, the dequantization process is expressed as:

[0045]

[0046] in, It is the dequantized feature data.

[0047] As a preferred embodiment of the method for visual laser data feature compression and extraction of the present invention, the step of restoring representative features also includes:

[0048] The tiled feature image is obtained through the DCT coefficient matrix, which is specifically expressed as:

[0049]

[0050] By decomposing the feature data into feature channels, the features removed by the encoder will be restored by S s The matching feature reconstruction in is specifically expressed as:

[0051]

[0052] in, For the representative characteristics of recovery, The features removed during the encoding process are then reconstructed, and the recovered features are transferred to the remaining network to continue the data analysis task.

[0053] In a second aspect, an embodiment of the present invention provides a visual laser data feature compression and extraction system, which includes a feature compression framework, a feature statistical analysis module, a feature quantification module, and a matching module;

[0054] The feature compression framework is based on a feature compression framework of a convolutional neural network and is used to extract intermediate features;

[0055] The feature statistical analysis module is used to determine redundant or low-information non-informative features and delete the features that are most similar to the non-informative features;

[0056] The feature quantization module is used to further compress statistical redundancy and eliminate redundancy between feature channels;

[0057] The matching module is used to determine the features with the least relevance to the monitoring task to form a non-informative subset, and select the features most similar to the non-informative features to form a similarity subset.

[0058] In a third aspect, an embodiment of the present invention provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the processor executes the computer program, it implements any step of the above-mentioned visual laser data feature compression and extraction method.

[0059] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, any step of the above-mentioned visual laser data feature compression and extraction method is implemented.

[0060] The beneficial effects of the present invention are: it can effectively compress intermediate features. The feature matching based on similarity is studied, the least important features in the monitoring task are determined, the features most relevant to the non-informative features are selected, and then the total features are divided into non-informative subsets, informative subsets and similarity subsets. The features in the similarity subset will be deleted in the encoder and restored according to the non-informative subset in the decoder, which significantly reduces the pressure of data transmission and storage, while ensuring the integrity and validity of the feature information. In addition, the present invention introduces a two-dimensional discrete cosine transform to further compress the redundancy of features, so that the feature information is flattened and forwarded to the DCT module to concentrate the feature values, significantly reducing the bit rate. The application of this technology not only enhances the real-time monitoring capability of the digital twin model for the equipment status, but also provides more accurate and sophisticated data support and basis for operation and maintenance decisions, thereby laying a solid technical guarantee for the efficient, stable and safe operation of power transmission and transformation equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing 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. Among them:

[0062] Figure 1 Flowchart of the feature compression and extraction method for visual laser data.

[0063] Figure 2 This is a structural diagram of the visual laser data feature compression and extraction method. DETAILED DESCRIPTION

[0064] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0065] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0066] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0067] The present invention is described in detail with reference to schematic diagrams. When describing the embodiments of the present invention, for the sake of convenience, the cross-sectional diagrams showing the device structure will not be partially enlarged according to the general scale, and the schematic diagrams are only examples, which should not limit the scope of protection of the present invention. In addition, in actual production, the three-dimensional dimensions of length, width and depth should be included.

[0068] At the same time, in the description of the present invention, it should be noted that the directions or positional relationships indicated by the terms "upper, lower, inner and outer" are based on the directions or positional relationships shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore cannot be understood as limiting the present invention. In addition, the terms "first, second or third" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0069] In the present invention, unless otherwise clearly specified and limited, the terms "install, connect, connect" should be understood in a broad sense, for example: it can be a fixed connection, a detachable connection or an integral connection; it can also be a mechanical connection, an electrical connection or a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0070] Example 1

[0071] Reference Figure 1 and Figure 2 , which is the first embodiment of the present invention, provides a method for visual laser data feature compression and extraction, comprising the following steps:

[0072] S1. Establish a feature compression framework based on convolutional neural network to extract intermediate features.

[0073] The steps of extracting intermediate features and determining phase velocity include:

[0074] The intermediate features include: non-information features, target information features and similarity feature subsets;

[0075] The selection of non-information features can be determined based on mutual information, correlation, statistics, etc. Since features with small variance generally contain fewer details and have less impact on the task, feature variance is used as the criterion. The computational complexity of variance is also affordable. Features are sorted in ascending order according to variance. The top-ranked features are selected as non-information features. The value of is related to the network layer of the feature and the monitoring task;

[0076] The subset of non-informative features is denoted as: S n ={f(l,i dx )|i dx ∈1,K};

[0077] Among them, i dx is the index of the non-informative feature channel;

[0078] Therefore, the similarity of non-informative features is expressed as:

[0079]

[0080] Among them, ρ i,j represents the similarity between f(l,i) and f(l,j), ρ i,j Determined by Pearson correlation coefficient, mutual information or l2 norm;

[0081] ρ i,j =||f(i,j)-f(l,j)||2.

[0082] The number of features is large, and calculating the entire similarity matrix is ​​very time-consuming. To simplify similarity matching, only the similarity between non-informative features and other features is calculated. That is, f(l,j)∈S n ,f(l,j)∈SS n ,i≠j, similarity matrix It can be simplified to K×(C–K). For any feature f(l,i dx )∈S n , matching their similar features can be expressed as an optimization problem:

[0083]

[0084] where j * is the same as f(l,i dx ) The index of the matching feature with the greatest similarity. In this way, S n The similar features of each feature in form a similarity subset S s .

[0085] S s ={f(l,j * )|j * ∈[1,C]}

[0086] Therefore S n and S s It is a one-to-one relationship and does not belong to S n and S s The remaining features of are assigned to the subset S i To reduce the amount of data, S will be removed from the output features. s , in the decoder, these deleted features will be passed to S n Finally, the features of the first layer are flattened to form a feature image x f (l)={f′(l,i|f′(l,i)∈S i ∪S s )}.

[0087] S2. According to the statistical characteristics of the features, determine the redundant or less informative uninformative features, and delete the features that are most similar to the uninformative features.

[0088] The discrete cosine transform is expressed as,

[0089] z f (l) = A T f′(l)A;

[0090] In the formula, z f (l) is the transformed feature data, A is the transformation matrix, [·] T is the matrix transpose.

[0091] Discrete cosine transform brings additional complexity. Generally speaking, the complexity increases with the size of the block. Small block size is generally less complex, but it is easy to cause more boundary jumps. This patent is based on the entire feature data for transformation. After the transformation, the data is more concentrated, which is conducive to subsequent entropy coding, so the total complexity does not increase significantly.

[0092] S3. Discrete cosine transform is introduced to further compress statistical redundancy, and redundancy between feature channels is eliminated through feature rearrangement, feature transformation and feature quantization.

[0093] The steps to eliminate redundancy between feature channels include:

[0094] Without loss of generality, uniform quantization is used and n-bit quantization is performed as follows:

[0095]

[0096] In the formula, max(z f (l)) and min(z f (l)) are the maximum and minimum values ​​in the feature data, [z q (l)] H,W is the quantized feature data, Indicates rounding down to the nearest integer. q (l) will be input into the arithmetic coding engine to further eliminate its statistical redundancy.

[0097] S4. In the receiver, the removed features are restored based on the features of the non-informative feature subset.

[0098] The steps to restore representative features include,

[0099] The feature extraction decoder includes: entropy decoding, inverse quantization, inverse transformation and feature recovery;

[0100] Among them, the dequantization process is expressed as:

[0101]

[0102] in, It is the dequantized feature data.

[0103] The steps of recovering representative features also include,

[0104] The tiled feature image is obtained through the DCT coefficient matrix, which is specifically expressed as:

[0105]

[0106] By decomposing the feature data into feature channels, the features removed by the encoder will be restored by S s The matching feature reconstruction in is specifically expressed as:

[0107]

[0108] in, For the representative characteristics of recovery, The features removed during the encoding process are then reconstructed, and the recovered features are transferred to the remaining network to continue the data analysis task.

[0109] The specific steps also include:

[0110] In a non-generative feedforward network, consecutive layers form a Markov chain. When the network weights are fixed, the output features of any layer are uniquely determined by the input through feedforward processing. The convolutional layer has multiple convolution kernels that can generate corresponding feature channels.

[0111] The input image is set to x, and the features obtained from the N middle layers of the network form a feature tensor space whose size is set to H×C×W, where C, H, and W represent the number of channels, height, and width of the feature, respectively. Before the nonlinear activation function, the output feature f(l) of the lth layer is given by the formula, which is expressed as:

[0112] f(l)=N(x,l)=∑h(l,i)*f(l-1,i)+α(l,i);

[0113] Where h(l,i) represents the convolution parameter of the i-th channel of the l-th layer, α(l,i) is the bias parameter of the convolutional neural network, and * represents the 2D convolution operation;

[0114] Determine the features with the least relevance to the monitoring task to form a non-informative subset S n , according to the similarity metric, select the features that are most similar to the non-informative features to form a similarity subset S s , S s and S nThe features in have a one-to-one correspondence, and then all the features are divided into three subsets;

[0115] S=S n ∪S i ∪S s ={f(l,i),i∈[1,C]};

[0116] In the formula, S represents the entire feature, S i is the set of remaining features. S s The features in will be removed in the encoder;

[0117] Output feature S ψ It is expressed as:

[0118] S ψ =Ψ(S);

[0119] In the receiver, the S n The corresponding features in restore the removed features;

[0120] S ψ The transformation is performed by transforming Γ[·]. Finally, the discrete code y(l) is generated by quantization and entropy coding, which is represented by C[·]:

[0121] y(l)=C(Γ(S ψ ),θ C );

[0122] In the formula, θ C Indicates the encoder parameter set to be tuned;

[0123] In the decoder, reconstruct the features:

[0124]

[0125] In the formula, It is the inverse process of the corresponding process in the encoder.

[0126] In summary, the intermediate features can be effectively compressed. The similarity-based feature matching is studied to determine the least important features in the monitoring task, select the features most relevant to the non-informative features, and then divide the total features into non-informative subsets, informative subsets, and similarity subsets. The features in the similarity subset will be deleted in the encoder and restored according to the non-informative subset in the decoder, which significantly reduces the pressure of data transmission and storage, while ensuring the integrity and validity of the feature information. In addition, the present invention introduces a two-dimensional discrete cosine transform to further compress the redundancy of features, so that the feature information is flattened and forwarded to the DCT module to concentrate the feature values, significantly reducing the bit rate. The application of this technology not only enhances the real-time monitoring capability of the digital twin model for the equipment status, but also provides more accurate and sophisticated data support and basis for operation and maintenance decisions, thereby laying a solid technical guarantee for the efficient, stable and safe operation of power transmission and transformation equipment.

[0127] Example 2

[0128] On the basis of the first embodiment, this embodiment further provides a visual laser data feature compression and extraction system, including a feature compression framework, a feature statistical analysis module, a feature quantification module, and a matching module;

[0129] The feature compression framework is based on a feature compression framework of a convolutional neural network and is used to extract intermediate features;

[0130] The feature statistical analysis module is used to determine redundant or low-information non-informative features and delete the features that are most similar to the non-informative features;

[0131] The feature quantization module is used to further compress statistical redundancy and eliminate redundancy between feature channels;

[0132] The matching module is used to determine the features with the least relevance to the monitoring task to form a non-informative subset, and select the features most similar to the non-informative features to form a similarity subset.

[0133] This embodiment also provides a computer device suitable for the visual laser data feature compression and extraction method, including a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute computer executable instructions to implement the visual laser data feature compression and extraction method proposed in the above embodiment.

[0134] The computer device may be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device may be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device may be a touch layer covering the display screen, or a key, trackball or touchpad provided on the housing of the computer device, or an external keyboard, touchpad or mouse, etc.

[0135] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, the method for compressing and extracting visual laser data features as proposed in the above embodiment is implemented.

[0136] The storage medium proposed in this embodiment and the data storage method proposed in the above embodiment belong to the same inventive concept. The technical details not fully described in this embodiment can be found in the above embodiment, and this embodiment has the same beneficial effects as the above embodiment.

[0137] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for compressing and extracting features of visual laser data, characterized in that: The following steps are included: Establish a feature compression framework based on convolutional neural networks to extract intermediate features; According to the statistical characteristics of the features, redundant or low-information uninformative features are determined, and the features that are most similar to the uninformative features are deleted; Discrete cosine transform is introduced to further compress statistical redundancy, and redundancy between feature channels is eliminated through feature rearrangement, feature transformation and feature quantization; In the receiver, the removed features are restored based on the features of the non-informative feature subset.

2. The method for visual laser data feature compression and extraction according to claim 1, characterized in that: The specific steps also include: The input image is set to x, and the features obtained from the N middle layers of the network form a feature tensor space whose size is set to H×C×W, where C, H, and W represent the number of channels, height, and width of the feature, respectively. Before the nonlinear activation function, the output feature f(l) of the lth layer is given by the formula, which is expressed as: f(l)=N(x,l)=∑h(l,i)*f(l-1,i)+α(l,i); Where h(l, i) represents the convolution parameter of the i-th channel of the l-th layer, α(l, i) is the bias parameter of the convolutional neural network, and * represents the 2D convolution operation; Determine the features with the least relevance to the monitoring task to form a non-informative subset S n , according to the similarity metric, select the features that are most similar to the non-informative features to form a similarity subset S s , S s and S n The features in have a one-to-one correspondence, and then all the features are divided into three subsets; S=S n ∪S i ∪S s ={f(l,i),i∈[1,C]}; In the formula, S represents the entire feature, S i is the set of remaining features. S s The features in will be removed in the encoder; Output characteristics It is expressed as: In the receiver, the S n The corresponding features in restore the removed features; The transformation is performed by transforming Γ[·]. Finally, the discrete code y(l) is generated by quantization and entropy coding, which is represented by C[·]: In the formula, θ C Indicates the encoder parameter set to be tuned; In the decoder, reconstruct the features: In the formula, It is the inverse process of the corresponding process in the encoder.

3. The method for visual laser data feature compression and extraction as claimed in claim 2, characterized in that: The steps of extracting intermediate features and determining phase velocity include: The intermediate features include: non-information features, target information features and similarity feature subsets; The subset of non-informative features is denoted as: S n ={f(l,i dx )|i dx ∈1,K}; Among them, i dx is the index of the non-informative feature channel; Therefore, the similarity of non-informative features is expressed as: Among them, ρ i，j represents the similarity between f(l, i) and f(l, j), ρ i，j Determined by Pearson correlation coefficient, mutual information or l2 norm; r i，j =||f(i,j)-f(l,j)||2.

4. The method for visual laser data feature compression and extraction as claimed in claim 3, characterized in that: The discrete cosine transform is expressed as, z f (l)=A T f′(l)A; In the formula, z f (l) is the transformed feature data, A is the transformation matrix, [·] T is the matrix transpose.

5. The method for visual laser data feature compression and extraction according to claim 4, characterized in that: The steps to eliminate redundancy between feature channels include: Without loss of generality, uniform quantization is used and n-bit quantization is performed as follows: In the formula, max(z f (l)) and min(z f (l)) are the maximum and minimum values ​​in the feature data, [z q (l)] H，W is the quantized feature data, Indicates rounding down to the nearest integer. q (l) will be input into the arithmetic coding engine to further eliminate its statistical redundancy.

6. The method for visual laser data feature compression and extraction according to claim 5, characterized in that: The steps to restore representative features include, feature The extraction decoder includes: entropy decoding, inverse quantization, inverse transformation and feature recovery; Among them, the dequantization process is expressed as: in, It is the dequantized feature data.

7. The method for visual laser data feature compression and extraction according to claim 6, characterized in that: The steps of recovering representative features also include, The tiled feature image is obtained through the DCT coefficient matrix, which is specifically expressed as: By decomposing the feature data into feature channels, the features removed by the encoder will be restored by S s The matching feature reconstruction in is specifically expressed as: in, For the representative characteristics of recovery, The features removed during the encoding process are then reconstructed, and the recovered features are transferred to the remaining network to continue the data analysis task.

8. A visual laser data feature compression and extraction system, based on the visual laser data feature compression and extraction method according to any one of claims 1 to 7, characterized in that: It includes feature compression framework, feature statistical analysis module, feature quantification module, and matching module; The feature compression framework is based on a feature compression framework of a convolutional neural network and is used to extract intermediate features; The feature statistical analysis module is used to determine redundant or low-information non-informative features and delete the features that are most similar to the non-informative features; The feature quantization module is used to further compress statistical redundancy and eliminate redundancy between feature channels; The matching module is used to determine the features with the least relevance to the monitoring task to form a non-informative subset, and select the features most similar to the non-informative features to form a similarity subset.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the visual laser data feature compression and extraction method described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the visual laser data feature compression and extraction method described in any one of claims 1 to 7 are implemented.

Images