Image compression reconstruction method and system based on SDA and cnn

Abstract

The present invention is suitable for the technical field of machine vision, and provides an image compression reconstruction method and system based on an SDA and a CNN. The image compression reconstruction method comprises the following steps: S1. rearranging original signals into a column by means of a sampling network to form a high-dimensional original signal, and sampling the high-dimensional original signal to obtain a measurement signal, wherein the original signals comprise original image data; S2. performing pixel enhancement and initial reconstruction on the measurement signal by means of a three-layer SDA neural network to obtain an initially reconstructed image; and S3. performing deep reconstruction processing on the original image data and the initially reconstructed image by means of a deep reconstruction network to obtain an optimal reconstructed image block. The present invention can improve the dimension and the reconstruction quality; in addition, the proposed framework is faster and more effective, and a reconstruction result better than that of a deep learning-based method of other similar structures is obtained at a low sampling rate.

Description

An image compression and reconstruction method and system based on SDA and CNN

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims priority to Chinese patent application No. 2023114280168, filed with the Patent Office of China on October 30, 2023, entitled "Image Compression and Reconstruction Method and System Based on SDA and CNN," the entire contents of which are incorporated herein by reference. Technical Field

[0003] The present invention relates to the field of machine vision technology, and in particular to an image compression and reconstruction method and system based on SDA and CNN. Background Art

[0004] Images are ubiquitous in the era of big data, and image processing technology has been developed and applied in many fields. However, in medical imaging, remote sensing, and other fields, image acquisition faces several limitations, such as limited data for describing features, massively parallel processing, the abundance of redundant information in cloud storage, and limited bandwidth for information transmission and storage space. Fortunately, the emergence of compressed sensing (CS) theory has successfully addressed these limitations.

[0005] Existing CS theory obtains sampling information at frequencies far below the Nyquist sampling frequency and recovers the original signal from the sampled information with high probability, breaking through the limitations of traditional sampling theorem. It allows people to use less information to represent more data without being affected by signal bandwidth and storage. However, traditional compressed sensing reconstruction algorithms struggle to reconstruct images in real time. Fortunately, with the widespread application of deep learning technology, deep learning-based reconstruction algorithms are orders of magnitude faster than traditional reconstruction algorithms. However, CS research is still constrained by the following two major challenges:

[0006] (1) First, traditional sampling and reconstruction methods rely on the prior assumption of signal sparsity. However, real data is not completely sparse in the transform domain. Therefore, how to learn the complex signal structure of real signals to improve reconstruction performance remains to be solved.

[0007] (2) Another is that at low measurement rates, the image quality reconstructed by traditional methods is poor and cannot even be used for image understanding tasks.

[0008] Summary of the Invention

[0009] The present invention provides an image compression and reconstruction method based on SDA and CNN, aiming to solve the problems of poor quality and low performance of existing image sampling and reconstruction.

[0010] In a first aspect, an embodiment of the present invention provides an image compression and reconstruction method based on SDA and CNN, the image compression and reconstruction method comprising the following steps:

[0011] S1. Rearranging the original signal into a column through a sampling network to form a high-dimensional original signal, and sampling the high-dimensional original signal to obtain a measurement signal; wherein the original signal includes original image data;

[0012] S2. Perform pixel enhancement and initial reconstruction on the measurement signal through a three-layer SDA neural network to obtain an initial reconstructed image;

[0013] S3. Perform deep reconstruction processing on the original image data and the initial reconstructed image through a deep reconstruction network to obtain an optimal reconstructed image block.

[0014] Preferably, in S1, the original signal includes a 33*33 image block.

[0015] Preferably, in S1, the expression of the measurement signal is as follows (1): y=T(W i x+b1)…(1);

[0016] The sampling network is defined as the first layer, W1 and b1 represent the weight and bias of the sampling network respectively, T represents the activation function ReLU, x∈R N Represents the input original signal, y∈R N Indicates the measurement signal.

[0017] Preferably, in said S2, said three-layer SDA neural network is respectively the second layer, the third layer and the fourth layer;

[0018] The measured signal is taken as input through the reconstruction network, and the initial reconstruction signal is output. The expression of each layer of the reconstruction network is (2): i =T(W i x i―1 +b i )…(2);

[0019] Among them, x i represents the output of layer i = 2, 3, 4, x1 specifically refers to the measurement signal and x1 = y; W i is the weight matrix of layer i, b i is the bias of the i-th layer, and T represents the nonlinear activation function ReLU.

[0020] Preferably, in S3, the deep reconstruction network is a convolutional neural network with 6 convolutional layers, and the parameters of the first three layers are the same as those of the last three layers.

[0021] Preferably, the sizes of the 6 convolutional layers are set to 11*11, 1*1, 7*7, 11*11, 1*1, 7*7 to keep the output feature map size still 33*33 image blocks, and the number of feature maps output in each layer is 64, 32, 1, 64, 32, 1.

[0022] Preferably, in S3, the optimal reconstructed image block is obtained by Adam and SGD optimization algorithms;

[0023] The Adam and SGD optimization algorithms are represented by a loss function, the loss function (3):

[0024] Among them, W L Expressed as all parameters, all parameters W L ={W i ,b i};W i and b i Represent the weight and bias of each layer, T(x i ,W L ) represents the output of the network.

[0025] In a second aspect, an embodiment of the present invention provides an image compression and reconstruction system based on SDA and CNN, the image compression and reconstruction system comprising:

[0026] a sampling module, configured to rearrange the original signal into a column through a sampling network to form a high-dimensional original signal, and obtain a measurement signal after sampling the high-dimensional original signal; wherein the original signal includes original image data;

[0027] An initial reconstruction module is used to perform pixel enhancement and initial reconstruction on the measurement signal through a three-layer SDA neural network to obtain an initial reconstructed image;

[0028] The deep reconstruction module is used to perform deep reconstruction processing on the original image data and the initial reconstructed image through a deep reconstruction network to obtain an optimal reconstructed image block.

[0029] Preferably, the original signal of the sampling module includes a 33*33 image block.

[0030] Compared with the existing technology, the present invention has the following advantages: the original signal is rearranged into a column through a sampling network to form a high-dimensional original signal, and the high-dimensional original signal is sampled to obtain a measurement signal; wherein the original signal includes original image data; the measurement signal is pixel-enhanced and initially reconstructed through a three-layer SDA neural network to obtain an initial reconstructed image; the original image data and the initial reconstructed image are deeply reconstructed through a deep reconstruction network to obtain an optimal reconstructed image block; this new framework jointly samples and reconstructs images by cascading SDA and CNN, in which SDA increases the dimensionality and preliminarily reconstructs the signal, and CNN further improves the reconstruction quality. Experiments show that compared with traditional iterative algorithms, the proposed framework is faster and more efficient, and achieves better reconstruction results at low sampling rates than other similar deep learning-based methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The present invention will be described in detail below with reference to the accompanying drawings. The above and other aspects of the present invention will become clearer and easier to understand through the detailed description made with reference to the following drawings. In the accompanying drawings:

[0032] FIG1 is a flow chart of an image compression and reconstruction method based on SDA and CNN provided by an embodiment of the present invention;

[0033] FIG2 is a framework diagram of an image compression and reconstruction method based on SDA and CNN provided in an embodiment of the present invention;

[0034] FIG3 is a module diagram of an image compression and reconstruction system based on SDA and CNN provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0035] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0036] Example 1

[0037] 1 and 2 , an embodiment of the present invention provides an image compression and reconstruction method based on SDA and CNN, the image compression and reconstruction method comprising the following steps:

[0038] S1. Rearranging the original signal into a column through a sampling network to form a high-dimensional original signal, sampling the high-dimensional original signal to obtain a measurement signal; wherein the original signal includes original image data. The original image data includes an input original picture or image block.

[0039] Specifically, the sampling and reconstruction of CS is considered a data-driven, end-to-end signal encoding and decoding process, without considering sparsity as a priori knowledge. Therefore, structural features are adaptively sampled and learned using the first layer of SDA, the fully connected layer. The sampling network is the first layer of this network, where neurons are connected to the original signal. The sampling network adaptively acquires key information required for reconstruction, and the sampling and reconstruction networks are trained jointly. The number of neurons in this fully connected layer varies with the measurement rate.

[0040] S2. Perform pixel enhancement and initial reconstruction on the measurement signal through a three-layer SDA neural network to obtain an initial reconstructed image.

[0041] Specifically, the method involves finding a mapping from low-dimensional to high-dimensional pixels for pixel enhancement and initial reconstruction, and employs a three-layer SDA neural network to achieve this goal. SDA has excellent performance in handling signal dimensionality issues. It can initially reconstruct the original image while simultaneously increasing the signal dimensionality.

[0042] S3. Perform deep reconstruction processing on the original image data and the initial reconstructed image through a deep reconstruction network to obtain an optimal reconstructed image block.

[0043] Since the initial reconstructed image blocks are still very different from the original image blocks, the gap between them needs to be further narrowed. Therefore, a convolutional neural network consisting of 6 convolutional layers is designed to achieve this goal. In order to maintain the size of the feature map, the pooling downsampling operation is abandoned and the zero padding operation is appropriately used.

[0044] Specifically, the original signal is rearranged into a column through a sampling network to form a high-dimensional original signal. After sampling the high-dimensional original signal, a measurement signal is obtained; the original signal includes original image data; the measurement signal is pixel-enhanced and initially reconstructed using a three-layer SDA neural network to obtain an initial reconstructed image; the original image data and the initial reconstructed image are deeply reconstructed using a deep reconstruction network to obtain an optimal reconstructed image block. This new framework, which jointly samples and reconstructs images by cascading SDA and CNN, increases the dimensionality and initially reconstructs the signal, while CNN further improves the reconstruction quality. Experiments show that compared with traditional iterative algorithms, the proposed framework is faster and more efficient, and achieves better reconstruction results at low sampling rates than other similar deep learning-based methods.

[0045] In this embodiment, in S1, the original signal includes a 33*33 image block, wherein the image block has 1089 image blocks, which facilitates rearranging the original signal into a column to form a high-dimensional original signal, and after sampling, obtains a corresponding measurement signal.

[0046] In this embodiment, in S1, the expression of the measurement signal is as follows (1): y=T(W i x+b1)…(1);

[0047] The sampling network is defined as the first layer, W1 and b1 represent the weight and bias of the sampling network respectively, T represents the activation function ReLU, x∈R N Represents the input original signal, y∈R N Indicates the measurement signal.

[0048] In this embodiment, in S2, the three-layer SDA neural network is the second layer, the third layer and the fourth layer respectively;

[0049] The measured signal is taken as input through the reconstruction network, and the initial reconstruction signal is output. The expression of each layer of the reconstruction network is (2): i =T(W i x i―1 +b i )…(2);

[0050] Among them, x i represents the output of layer i = 2, 3, 4, x1 specifically refers to the measurement signal and x1 = y; W i is the weight matrix of layer i, b i is the bias of the i-th layer, and T represents the nonlinear activation function ReLU.

[0051] Among them, x i The dimension of is related to the number of neurons in the SDA layer. Since the original image patch size is 33*33, we set the number of neurons in the three SDA layers to 1089, 272, and 1089, respectively. Then, the image patch is reorganized into the original image structure by x4.

[0052] In this embodiment, in S3, the deep reconstruction network is a convolutional neural network with six convolutional layers, with the first three layers having the same parameters as the last three layers. This goal is achieved by designing a convolutional neural network consisting of six convolutional layers. To maintain the size of the feature map, we discard the pooling downsampling operation and use zero padding as appropriate.

[0053] In this example, the sizes of the six convolutional layers are set to 11*11, 1*1, 7*7, 11*11, 1*1, and 7*7 to maintain the output feature map size of a 33*33 image block. The number of feature maps output at each layer is 64, 32, 1, 64, 32, and 1. Except for the last layer, the other five layers apply the nonlinear activation function ReLU. The last layer is the output layer of the network and outputs the reconstructed image block.

[0054] In this embodiment, in S3, the optimal reconstructed image block is obtained by Adam and SGD optimization algorithms;

[0055] The Adam and SGD optimization algorithms are represented by a loss function, the loss function (3):

[0056] Among them, W L Expressed as all parameters, all parameters W L ={W i ,b i};W i and b i Represent the weight and bias of each layer, T(x i ,W L ) represents the output of the network.

[0057] Specifically, to reduce parameters and balance the computational complexity of the network with reconstruction performance, we divide the original image into 33*33 sub-image blocks. The number of image blocks in the dataset varies, depending on the size of each image. The total number of image blocks in the dataset is S.

[0058] Through the joint training of the sampling layer and the reconstruction network, the reconstruction network can guide the optimization direction of the sampling layer and enable the sampling layer to adaptively obtain the information required for reconstruction.

[0059] Adam and SGD optimization algorithms are used to optimize all parameters W L ={W i ,b i}(We use two optimization algorithms to optimize the network separately and select the network with lower loss as the final sampling reconstruction network).

[0060] Among them, the Adam optimization algorithm is an optimization algorithm used to replace stochastic gradient descent in deep learning models. It combines the optimal performance of the AdaGrad and RMSProp algorithms. The parameter adjustment is relatively simple, and the default parameters can handle most problems.

[0061] The SGD optimization algorithm is a gradient-based optimization algorithm used to update the parameters of deep neural networks. Its basic idea is that, at each iteration, a mini-batch of samples is randomly selected to calculate the gradient of the loss function and use this gradient to update the parameters. This randomness makes the algorithm more robust, preventing it from falling into local minima and resulting in faster training.

[0062] Example 2

[0063] An embodiment of the present invention provides an image compression and reconstruction system 200 based on SDA and CNN, the image compression and reconstruction system 200 comprising:

[0064] A sampling module 201 is configured to rearrange the original signal into a column through a sampling network to form a high-dimensional original signal, and obtain a measurement signal after sampling the high-dimensional original signal; wherein the original signal includes original image data;

[0065] An initial reconstruction module 202 is configured to perform pixel enhancement and initial reconstruction on the measurement signal through a three-layer SDA neural network to obtain an initial reconstructed image;

[0066] The deep reconstruction module 203 is configured to perform deep reconstruction processing on the original image data and the initial reconstructed image through a deep reconstruction network to obtain an optimal reconstructed image block.

[0067] Specifically, the sampling module 201 is used to rearrange the original signal into a column through a sampling network to form a high-dimensional original signal. After sampling the high-dimensional original signal, a measurement signal is obtained; wherein, the original signal includes original image data; the initial reconstruction module 202 performs pixel enhancement and initial reconstruction on the measurement signal through a three-layer SDA neural network to obtain an initial reconstructed image; the deep reconstruction module 203 performs deep reconstruction processing on the original image data and the initial reconstructed image through a deep reconstruction network to obtain an optimal reconstructed image block. This new framework jointly samples and reconstructs images by cascading SDA and CNN. SDA increases the dimensionality and preliminarily reconstructs the signal, while CNN further improves the reconstruction quality. Experiments show that compared with traditional iterative algorithms, the proposed framework is faster and more efficient, and achieves better reconstruction results at low sampling rates than other deep learning-based methods with similar structures.

[0068] In this embodiment, the original signal of the sampling module includes a 33*33 image block, wherein the image block has 1089 image blocks, which facilitates rearranging the original signal into a column to form a high-dimensional original signal, and after sampling, obtains the corresponding measurement signal.

[0069] The image compression and reconstruction system 200 based on SDA and CNN can implement the steps in the image compression and reconstruction method based on SDA and CNN in the above embodiment, and can achieve the same technical effects. Please refer to the description in the above embodiment and will not repeat them here.

[0070] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or apparatus comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or apparatus comprising the element.

[0071] The embodiments of the present invention are described above in conjunction with the accompanying drawings. What is disclosed is only a preferred embodiment of the present invention. However, the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms and equivalent changes without departing from the scope of protection of the purpose of the present invention and the claims, which are all within the protection of the present invention.

Claims

1. An image compression and reconstruction method based on SDA and CNN, characterized in that: The image compression and reconstruction method comprises the following steps: S1. Rearranging the original signal into a column through a sampling network to form a high-dimensional original signal, and obtaining a measurement signal after sampling the high-dimensional original signal; wherein the original signal includes original image data; S2, performing pixel enhancement and initial reconstruction on the measurement signal through a 3-layer SDA neural network to obtain an initial reconstructed image; S3. Performing deep reconstruction processing on the original image data and the initial reconstructed image through a deep reconstruction network to obtain an optimal reconstructed image block.

2. The image compression and reconstruction method based on SDA and CNN as claimed in claim 1, characterized in that: In S1, the original signal includes a 33*33 image block.

3. The image compression and reconstruction method based on SDA and CNN as claimed in claim 2, characterized in that: In S1, the expression of the measurement signal is as follows (1): y = T (W i x+b1)…(1); Wherein, the sampling network is defined as the first layer, W1 and b1 represent the weight and bias of the sampling network respectively, T represents the activation function ReLU, x∈R N Represents the input original signal, y∈R N Indicates the measurement signal.

4. The image compression and reconstruction method based on SDA and CNN as claimed in claim 3, characterized in that: In S2, the three-layer SDA neural network is respectively the second layer, the third layer and the fourth layer; The measured signal is taken as input through the reconstruction network, and the initial reconstructed signal is output. The expression of each layer of the reconstruction network is (2): x i =T(W i x i―1 +b i )…(2); Among them, x i represents the output of the i-th layer i=2,3,4, x1 specifically refers to the measurement signal and x1=y; W i is the weight matrix of the i-th layer, b i is the bias of the i-th layer, and T represents the nonlinear activation function ReLU.

5. The image compression and reconstruction method based on SDA and CNN as claimed in claim 4, characterized in that: In S3, the deep reconstruction network is a convolutional neural network with 6 convolutional layers, and the parameters of the first three layers are the same as those of the last three layers.

6. The image compression and reconstruction method based on SDA and CNN as claimed in claim 5, characterized in that: The sizes of the 6 convolutional layers are set to 11*11, 1*1, 7*7, 11*11, 1*1, 7*7 to keep the output feature map size still 33*33 image blocks. The number of feature maps output in each layer is 64, 32, 1, 64, 32, 1.

7. The image compression and reconstruction method based on SDA and CNN as claimed in claim 6, characterized in that: In S3, the optimal reconstructed image block is obtained by using Adam and SGD optimization algorithms; The Adam and SGD optimization algorithms are represented by a loss function, the loss function (3): Among them, W L Expressed as all parameters, all parameters W L = {W i ,b i }; W i and b i Represent the weight and bias of each layer, T(x i ,W L ) represents the output of the network.

8. An image compression and reconstruction system based on SDA and CNN, characterized in that: The image compression and reconstruction system comprises: A sampling module, used for rearranging the original signal into a column through a sampling network to form a high-dimensional original signal, and obtaining a measurement signal after sampling the high-dimensional original signal; wherein the original signal includes original image data; An initial reconstruction module, used for performing pixel enhancement and initial reconstruction on the measurement signal through a three-layer SDA neural network to obtain an initial reconstructed image; The deep reconstruction module is used to perform deep reconstruction processing on the original image data and the initial reconstructed image through a deep reconstruction network to obtain an optimal reconstructed image block.

9. The image compression and reconstruction system based on SDA and CNN as claimed in claim 8, characterized in that: The original signal of the sampling module includes a 33*33 image block.

Images