A bit-level compression and reconstruction method for tensor data based on quantized neural network

Through the bit-level compression and reconstruction method based on quantized neural network, combined with the lightweight neural quantization encoder and channel fusion feature enhancement module, the problems of transmission bandwidth and reconstruction quality in high-dimensional tensor data compression and reconstruction are solved, and high efficiency is achieved Low bit rate compression and high precision reconstruction.

CN116824318BActive Publication Date: 2024-11-29CHONGQING UNIV
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202310547390.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-05-15
Publication Date
2024-11-29
Estimated Expiration
2043-05-15

AI Technical Summary

Technical Problem

The prior art has problems in high-dimensional tensor data compression and reconstruction, with high data throughput, limited transmission bandwidth and low reconstruction quality. The frequency domain transformation and compression perception methods have their own shortcomings, so it is impossible to effectively combine bit-level compression and Advantages of feature-level compression.

Method used

Using a bit-level compression and reconstruction method based on quantized neural networks, the combination of feature-level compression and bit-level compression is achieved through a lightweight neural quantization encoder and a super decoder based on channel fusion feature enhancement, thereby improving the expressive ability of compressed features. , and efficient compression and reconstruction of data through Hoffman encoding and channel fusion feature enhancement modules.

Benefits of technology

It realizes efficient compression and reconstruction of high-dimensional tensor data in edge devices with limited computing resources, significantly reduces the data transmission bandwidth requirement, improves reconstruction quality, maintains high accuracy at low bit rate, and shows good compression feature expression Capabilities and reconstruction effects.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN116824318B_ABST
    Figure CN116824318B_ABST
Patent Text Reader

Abstract

The present invention relates to a method for bit-level compression and reconstruction of tensor data based on a quantized neural network, which belongs to the field of deep learning and comprises a lightweight neural quantization encoder and a super-resolution decoder based on channel fusion feature enhancement; the lightweight neural quantization encoder realizes feature-level compression of an input tensor through a convolution operation, combines neural quantization output with Huffman coding to realize bit-level compression, and finally realizes two-level compression combining feature-level compression and bit-level compression; the super-resolution decoder based on channel fusion feature enhancement performs Huffman decoding on compressed data through a Huffman decoder, enhances the feature expression ability of compressed data through a channel fusion feature enhancement backbone network, completes two-dimensional spatial upsampling of tensor data through a spatial super-resolution module using linear interpolation, and finally restores the channel resolution of the original tensor data through a channel super-resolution convolution module to obtain a reconstructed tensor.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of deep learning and relates to a tensor data bit-level compression and reconstruction method based on a quantized neural network. Background Art

[0002] High-dimensional tensor data (such as video data and hyperspectral data) has the characteristics of large data volume and high throughput. There are three main challenges in its generation, transmission and processing: ① High data throughput and insufficient computing resources of data acquisition sensors; ② Extremely limited data transmission bandwidth and large compression requirements; ③ In order to smoothly carry out subsequent research and application, the reconstruction quality of compressed tensor data must meet high requirements. In response to the above challenges, the following three types of methods have been proposed:

[0003] 1. As a standard compression coding method in the industry, the frequency domain transform method divides the input high-dimensional tensor data into several sub-blocks, and then uses discrete cosine transform or discrete wavelet transform to perform transform coding, and only retains low-frequency coefficients as compression targets. On this basis, other technologies such as principal component analysis are combined to complete the compression task of tensor data. However, this method only focuses on bit-level compression and lacks specific operations to compensate for the decrease in reconstruction accuracy caused by compression.

[0004] 2. Compressed sensing-based methods assume that signal data has sparse or nearly sparse properties in certain specific domains, and that the signal data can be recovered from its compressed results. The reconstruction of compressed results is an ill-posed problem for compressed sensing, which has given rise to regularization methods such as sparsity-based and total variation-based. However, most of the above-mentioned compressed sensing-based methods have two disadvantages: first, the compression operation of the above-mentioned methods ignores the compression potential at the bit level, resulting in limited compression performance; second, the method has high requirements on the assumptions of sparsity, total variation or low rank of high-dimensional tensor data in the transform domain, but cannot fully capture the profound prior knowledge behind the data, which may lead to suboptimal reconstruction results.

[0005] 3. In recent years, the potential of deep neural network (DNN)-based methods in the field of tensor data compression and reconstruction has been continuously explored. This type of method uses end-to-end training to learn the compression and reconstruction capabilities of tensor data from a large amount of data. Compared with compressed sensing-based methods, DNN-based methods can directly learn potential related information from training samples, thus requiring less prior knowledge in tensor data; and compared with frequency domain transformation-based methods, DNN-based methods can achieve bit-level compression while compensating for the reduction in reconstruction accuracy caused by compression by introducing and optimizing learnable parameters. However, the compression strategies of most DNN-based methods only stay at the feature level, without combining the advantages of bit-level compression, resulting in insufficient compression rate; in addition, in the reconstruction stage of tensor data, these methods do not pay enough attention to the expression of compressed features, which may lead to low reconstruction quality. Summary of the invention

[0006] In view of this, the purpose of the present invention is to reduce the demand for data transmission bandwidth resources, improve the expression ability of compressed features, and provide a tensor data bit-level compression and reconstruction method based on quantized neural network.

[0007] In order to achieve the above object, the present invention provides the following technical solutions:

[0008] A method for bit-level compression and reconstruction of tensor data based on quantized neural network, comprising the following steps:

[0009] S1: Get tensor data, pre-process it and divide it into training set and test set;

[0010] S2: construct a bit-level tensor data compression and reconstruction neural network, use the training set for training, and use the test set for testing; the bit-level tensor data compression and reconstruction neural network includes a lightweight neural quantization encoder and a super-resolution decoder based on channel fusion feature enhancement; the lightweight neural quantization encoder realizes feature-level compression of the input tensor through convolution operation, combines the neural quantization output with Huffman coding to realize bit-level compression, and finally realizes two-level compression combining feature-level compression and bit-level compression; the super-resolution decoder based on channel fusion feature enhancement consists of a Huffman decoder, a channel fusion feature enhancement backbone network, and a space-channel super-resolution module; the compressed data is Huffman decoded by the Huffman decoder, the feature expression ability of the compressed data is improved by the channel fusion feature enhancement backbone network, and the space-channel super-resolution module uses linear interpolation and convolution operations to complete the spatial and channel upsampling of the tensor data, restore the spatial and channel resolution of the original tensor data, and obtain the reconstructed tensor;

[0011] S3: Compress and reconstruct the tensor data using the trained bit-level tensor data compression and reconstruction neural network.

[0012] Furthermore, in the lightweight neural quantization encoder, for a given input tensor data X, the output result Y is obtained after convolution neural quantization. q , defined as:

[0013]

[0014] Where W represents the convolution kernel weight of the corresponding convolution layer, and the function represents the convolution-GELU operation, β represents the GELU activation function, represents the convolution operation; Q() represents the neural quantization operation. For each full-precision element y, its value is first restricted to [0, α] by formula (2), where the upper limit α is learnable. Thus, the b-bit quantized value Q(y) is calculated as:

[0015]

[0016] Where rounding() is the rounding function.

[0017] This method performs neural quantization on both the activation and weight of the neural network. After that, the quantized tensor data Y q The result is input into the Huffman encoder for further compression coding. Since the convolution part belongs to a shallow network model and the convolution kernel weights are quantized, it is lightweight. The Huffman operation does not have any parameters and can be directly used as part of the hardware design, occupying less memory. Therefore, the neural quantization encoder can be successfully deployed in edge devices with less computing and storage resources.

[0018] Furthermore, the channel fusion feature enhancement backbone network is composed of n channel fusion feature enhancement modules, wherein each channel fusion feature enhancement module is composed of three channel feature fusion modules and one channel feature enhancement module;

[0019] The channel feature fusion module includes three convolution-SiLU layers, where SiLU is an activation function. For each convolution-SiLU layer, the feature maps of all previous layers are densely connected and stacked, and used as the input feature maps of this layer. Then, the sum of the input feature tensor and the output feature tensor of the last layer is realized through residual connection to achieve channel feature fusion.

[0020] The channel feature enhancement module first performs spatial global average pooling on the tensor data of the c channels input. avgpool and the maximum pooling f maxpool Get a 1×1×c feature sequence, and then use a group convolution with a convolution kernel size of 2×1 The global average pooling f avgpooland the maximum pooling f maxpool The feature sequence is fused, keeping the number of channels c unchanged, and finally input into the multilayer perceptron f MLP Get the corresponding channel fusion feature enhancement coefficient M FE , the expression is as follows:

[0021]

[0022] where Y T is the input feature tensor of the channel feature enhancement module, σ is the sigmoid activation function, and finally the channel fusion feature enhancement coefficient M is used FE Y T The weighted sum is then obtained through residual connection to obtain the output Y of the entire channel fusion feature enhancement module. CA :

[0023] Y CA =Y IN +Y T ⊙M FE (4)

[0024] where Y IN It is the input of the channel fusion feature enhancement module, and ⊙ represents the multiplication of the corresponding elements of the tensor.

[0025] The corresponding loss function is used to measure the accuracy of tensor data reconstruction, and the corresponding network weights are obtained by minimizing the loss function training and optimization.

[0026] Furthermore, a lightweight neural quantization encoder is deployed on the data acquisition device side, and a two-level compression operation is performed after data perception.

[0027] The beneficial effects of the present invention are:

[0028] 1. The encoder has a lightweight structure and can be easily deployed in edge devices with limited computing resources.

[0029] 2. The encoder adopts a two-level compression strategy that combines feature-level compression and bit-level compression to achieve low-bitrate compression of high-dimensional tensor data and minimize the impact of limited data transmission bandwidth.

[0030] 3. The decoder proposes a new module - channel fusion feature enhancement module as the basic unit of the backbone network, which improves the expression ability of compressed features and combines a series of super-resolution operations to reconstruct the original tensor data with higher accuracy.

[0031] 4. By compressing and reconstructing the RGB original images of the Moffett dataset using the encoder and decoder provided by the present invention, it is found that there is almost no obvious difference between the visualized images of each reconstructed data and the original images, and the texture, edge and other details in the reconstructed images can be clearly seen, showing good reconstruction quality.

[0032] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0034] Figure 1 The overall workflow diagram for bit-level tensor data compression and reconstruction of neural networks;

[0035] Figure 2 This is a diagram of the structure of a lightweight neural quantization encoder;

[0036] Figure 3 This is a structural diagram of a super-resolution decoder based on channel fusion feature enhancement;

[0037] Figure 4 This is the structure diagram of the channel fusion feature enhancement module;

[0038] Figure 5 The original RGB image of the Moffett dataset and the corresponding compressed RGB image of the reconstructed hyperspectral data, where (a) is the original image, (b) is the reconstructed image when the feature tensor is compressed with full precision (32 bits), (c) is the image under b=11, (d) is the image under b=10, (e) is the image under b=9, (f) is the image under b=8, and (g) is the image under b=7. DETAILED DESCRIPTION

[0039] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0040] Among them, the drawings are only used for illustrative explanations, and they only represent schematic diagrams rather than actual pictures, and should not be understood as limitations on the present invention. In order to better illustrate the embodiments of the present invention, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0041] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if the terms "upper", "lower", "left", "right", "front", "rear", etc. indicate the orientation or position relationship, they are based on the orientation or position relationship shown in the drawings, which is 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 orientation, be constructed and operate in a specific orientation. Therefore, the terms describing the position relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0042] See also Figure 1 to Figure 5 The present invention provides a method for bit-level compression and reconstruction of tensor data based on a quantized neural network. The method performs bit-level compression and reconstruction on tensor data based on a bit-level tensor data compression and reconstruction neural network. The network is mainly composed of a lightweight neural quantization encoder and a super-resolution decoder based on channel fusion feature enhancement. The lightweight neural quantization encoder can be deployed in edge devices for data acquisition. Its function is to achieve two-level compression (feature-level compression and bit-level compression) of tensor data, achieve a higher compression rate, and thus minimize the demand for data transmission bandwidth resources; the super-resolution decoder based on channel fusion feature enhancement adopts a channel fusion feature enhancement module to improve the expression ability of compressed features, and combines super-resolution technology to achieve high-precision reconstruction of original tensor data from compressed data.

[0043] In order to highlight the technical objectives, solutions and advantages of the present invention, the embodiments of the present invention take hyperspectral tensor data as an example and describe the implementation methods of the present invention in detail in combination with the accompanying drawings.

[0044] 1) Data Introduction

[0045] Hyperspectral images containing 224 frequency bands were collected using the hyperspectral sensor AVIRIS in thirteen regions in the United States and Canada, including California, Oregon, and Yellowstone National Park.

[0046] 2) Data processing

[0047] The hyperspectral image in 1) was preprocessed by frequency bands, and the water absorption bands and low-quality frequency bands were removed, with serial numbers 1-10, 104-116, 152-170, and 215-224. 172 frequency bands were retained, and the hyperspectral image after frequency band preprocessing was cropped to obtain 1183 hyperspectral sub-images with a size of 256×256×172. These sub-images were then mean normalized and divided into training set and test set in a ratio of 9:1.

[0048] 3) Compression encoding

[0049] like Figure 2 As shown in the figure, the lightweight neural quantization encoder adopts the "convolutional neural quantization + entropy coding" strategy, which realizes feature-level compression of the input tensor through convolution operation, combines the neural quantization output with Huffman coding to achieve bit-level compression, and finally realizes two-level compression combining feature-level compression and bit-level compression.

[0050] In this embodiment, each hyperspectral sub-image is cut into 128×4×172 strip data before compression encoding. Figure 1 , Figure 3 As shown, after obtaining the strip data, the lightweight neural quantization encoder first performs feature-level compression on it, which is completed by the convolutional neural quantization operation:

[0051]

[0052] Where W represents the convolution kernel weight of the corresponding convolution layer, and the function represents the convolution-GELU operation, β represents the GELU activation function, represents the convolution operation, Q() represents the neural quantization operation. For each full-precision element y, its value is first limited to [0, α] by the following formula, where the upper limit α is learnable, so that the b-bit quantization value Q(y) can be calculated:

[0053]

[0054] Where rounding() is a rounding function. This method performs neural quantization on the activation and weight of the neural network at the same time. After that, the quantized tensor data Yq is input into the Huffman encoder for further compression encoding, and the compressed bit stream obtained is finally transmitted to the receiving end. Since the convolution part belongs to a shallow network model and the convolution kernel weights are quantized, it is lightweight. The Huffman operation does not have any parameters and can be directly used as part of the hardware design, occupying less memory. Therefore, the neural quantization encoder can be smoothly deployed in edge devices with less computing and storage resources. The convolution process can achieve feature compression at a specific compression rate by setting parameters such as the number of convolution kernels, the convolution step size, and the amount of padding. In order to finally achieve low-bitrate compression encoding of hyperspectral tensor data, this embodiment sets the feature compression rate to 1%. The relevant parameters of the entire feature-level compression process are shown in Table 1.

[0055] Table 1

[0056] Layer No. Input Feature Kernel Size Stride Padding Output Feature 1 128×4×172 3×3 [2,2] [1,0] 64×1×128 2 64×1×128 3×1 [1,1] [1,0] 64×1×64 3 64×1×64 3×1 [2,1] [1,0] 32×1×27

[0057] In Table 1, Layer No. indicates the convolutional neural quantization layer number. In this embodiment, three layers of convolution are built. InputFeature indicates the input feature map size of this layer, Kernel Size indicates the convolution kernel size of this layer, Stride indicates the convolution step size of this layer, Padding indicates the padding amount of this layer, and Output Feature indicates the output feature map size of this layer. As shown in Table 1, the final output feature map size is 32×1×27, indicating that the features within and between channels of the input tensor data are compressed, and the compression rate is 1%.

[0058] 4) Decoding and reconstruction

[0059] like Figure 3 As shown in the figure, after the bit stream compressed and output by the lightweight neural quantization encoder is transmitted through the channel, it is received and decoded and reconstructed by the super-resolution decoder based on channel fusion feature enhancement. The decoder consists of three parts: a Huffman decoder, a channel fusion feature enhancement backbone network, and a spatial-channel super-resolution module. First, the compressed data is Huffman decoded to restore the quantized tensor data Yq, and then Yq is input into the channel fusion feature enhancement backbone network to improve the feature expression ability of the compressed data. The backbone network is mainly composed of 16 channel fusion feature enhancement modules, each of which is composed of three channel feature fusion modules and one channel feature enhancement module. As shown in Figure 4As shown in the figure (upper right), each channel feature fusion module combines the characteristics of ResNet and DenseNet structures, including 3 convolution-SiLU layers, where SiLU is the activation function. For each convolution-SiLU layer, all feature maps of the previous layers are densely connected and used as the input feature map of this layer (DenseNet structure), and then the sum of the input feature tensor and the output feature tensor of the last layer is realized through residual connection (ResNet structure), so as to achieve the purpose of channel feature fusion. The structure of the channel feature enhancement module is shown in Figure 1. Figure 4 As shown in (lower right), for the input tensor data of c channels, the global average pooling f is first performed in space avgpool and the maximum pooling f maxpool Get a 1×1×c feature sequence, and then use a group convolution with a convolution kernel size of 2×1 The above two feature sequences are fused, keeping the number of channels c unchanged, and finally input into the multilayer perceptron f MLP Get the corresponding channel fusion feature enhancement coefficient M FE , the expression is as follows:

[0060]

[0061] where Y T is the input feature tensor of the channel feature enhancement module, σ is the sigmoid activation function. Finally, the channel fusion feature enhancement coefficient M is used FE Y T The weighted sum is then obtained through residual connection to obtain the output Y of the entire channel fusion feature enhancement module. CA :

[0062] Y CA =Y IN +Y T ⊙M FE (4)

[0063] where Y IN It is the input of the channel fusion feature enhancement module, and ⊙ represents the multiplication of the corresponding elements of the tensor. After the compressed features are enhanced by the backbone network, the spatial-channel super-resolution module uses linear interpolation and convolution operations to complete the spatial and channel upsampling of the tensor data, restore the 172 frequency bands of the original hyperspectral data, and obtain the reconstructed tensor.

[0064] 5) Model optimization and testing

[0065] Since the L1 norm can keenly capture the reconstructed tensor data The small error between the original tensor data X is used to measure the reconstructed tensor data The gap between the original tensor data X and the original tensor data X is used as the target loss function of the model during training and optimization:

[0066]

[0067] Where B is the batch size of training data, and P represents the number of two-dimensional space elements of tensor data. After setting the learning rate η and batch size B, the Adam algorithm is used as the optimizer for model learning, and then the model is trained on the training set. When the loss function converges, the training ends, the corresponding model weight parameters are obtained, and various performance indicators are tested on the corresponding test set.

[0068] The experimental results are shown in Table 2, which shows the performance of the model when the quantization bit number b is 7-11. The lower the quantization bit number of the compressed feature tensor, the greater the degree of compression. As shown in the table, when the quantization bit number b of the neural quantizer is 7, the Huffman compression rate reaches 63.9%, and the compression bit rate reaches an extremely low level of 0.045bpppb. At the same time, the model also shows good reconstruction performance. When the bit rate is lower than 0.1bpppb, the reconstruction PSNR exceeds 34dB. The PSNR under 10-bit and 11-bit quantization of the compressed feature tensor is even better than the performance of full precision (32bit).

[0069] Table 2

[0070]

[0071] Figure 5 The original RGB image of the Moffett dataset and the corresponding RGB image of the compressed reconstructed hyperspectral data are given. (a) is the original image, (b) is the reconstructed image when the feature tensor is compressed with full precision (32bit), b=11 in (c), b=10 in (d), b=9 in (e), b=8 in (f), and b=7 in (g). It can be observed that there is almost no obvious difference between the visualized images of each reconstructed data and the original images, and the texture, edge and other details in the reconstructed image can be clearly seen, showing good reconstruction quality.

[0072] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A method for bit-level compression and reconstruction of tensor data based on quantized neural network, characterized in that: The following steps are involved: S1: Collect hyperspectral images using hyperspectral sensors; Perform frequency band preprocessing on the hyperspectral image to remove water absorption bands and low-quality frequency bands; The hyperspectral image after frequency band preprocessing is cropped to obtain a hyperspectral sub-image, and then the hyperspectral sub-image is mean normalized, and a training set and a test set are divided in proportion; S2: constructing a bit-level tensor data compression and reconstruction neural network, using the training set for training, and using the test set for testing; the bit-level tensor data compression and reconstruction neural network includes a lightweight neural quantization encoder and a super-resolution decoder based on channel fusion feature enhancement; The lightweight neural quantization encoder realizes feature-level compression of hyperspectral tensor data through convolution operation, combines neural quantization output with Huffman coding to realize bit-level compression, and finally realizes two-level compression combining feature-level compression and bit-level compression; the super-resolution decoder based on channel fusion feature enhancement is composed of a Huffman decoder, a channel fusion feature enhancement backbone network, and a space-channel super-resolution module; the compressed data is Huffman decoded by the Huffman decoder, the feature expression ability of the compressed data is improved by the channel fusion feature enhancement backbone network, and the space-channel super-resolution module uses linear interpolation and convolution operations to complete the spatial and channel upsampling of the hyperspectral tensor data, restore the spatial and channel resolution of the original hyperspectral tensor data, and obtain a reconstructed tensor; S3: compress and reconstruct the hyperspectral tensor data using the trained bit-level tensor data compression and reconstruction neural network; The channel fusion feature enhancement backbone network is composed of n The module is composed of three channel feature fusion modules and one channel feature enhancement module. The channel feature fusion module includes three convolution-SiLU layers, where SiLU is an activation function. For each convolution-SiLU layer, the feature maps of all previous layers are densely connected and stacked, and used as the input feature map of this layer. Then, the sum of the input feature tensor and the output feature tensor of the last layer is realized through residual connection to realize channel feature fusion. The channel feature enhancement module is used for the input The tensor data of channels is first pooled globally on the space and max pooling get The feature sequence is then convolution kernel size is Grouped convolution The global average pooling and max pooling Feature sequence fusion, maintaining the number of channels unchanged, and finally input into the multilayer perceptron Get the corresponding channel fusion feature enhancement coefficient , the expression is as follows: (3) in is the input feature tensor of the channel feature enhancement module, is the sigmoid activation function, and finally the channel fusion feature enhancement coefficient is used right The weighted sum is then obtained through residual connection to obtain the output of the entire channel fusion feature enhancement module. : (4) in is the input of the channel fusion feature enhancement module, Represents the multiplication of corresponding elements of tensors.

2. The method for bit-level compression and reconstruction of tensor data based on quantized neural network according to claim 1, characterized in that: In the lightweight neural quantization encoder, each hyperspectral sub-image is cropped into strip data as input tensor data , the output result is obtained after convolutional neural quantization , defined as: (1) in Represents the convolution kernel weight of the corresponding convolution layer, function represents the convolution-GELU operation, represents the GELU activation function, Represents the convolution operation; Represents a neural quantization operation, for each full-precision element , first limit its value to , where the upper bound value It is learnable, so Bit quantization value By calculation, we get: (2) in is the rounding function.

3. The method for bit-level compression and reconstruction of tensor data based on quantized neural network according to claim 1, characterized in that: The lightweight neural quantization encoder is deployed on the data acquisition device side, and two-level compression operations are performed after data perception.

Citation Information

Patent Citations

  • Super-resolution image reconstruction method based on lightweight network

    CN110009565A

  • Compressed video quality enhancement method based on reconstructed flow field

    CN116012272A