Lattice-reduction-based estimation of color space for jpeg image compression

Abstract

The present application belongs to the technical field of image compression, and provides a JPEG image compression color space estimation method based on lattice reduction, mainly including the following steps: S1. selecting a color image and performing DCT transformation; S2. selecting a DCT frequency, and obtaining a 3D DCT coefficient vector histogram according to different 8x8 blocks; S3. descendingly sorting the obtained histogram; S4. selecting the first vector outside the range in the sorting list as a lattice base vector; S5. selecting the next vector, and if there is no remaining vector in the list, then exiting; S6. calculating the error between the closest vector on the lattice formed by the current selected vector and the current base vector; S7. judging the error vector, and when meeting the requirements, performing S5; and S8. obtaining the base vectors estimated from different DCT frequencies, and calculating the color transformation matrix according to the base vectors. The present application can estimate the unknown compression color space in the JPEG compression process through a given color image quantized in the compression color space.

Description

Technical Field

[0001] This invention proposes a color space estimation method for JPEG image compression based on lattice basis reduction, belonging to the field of image compression technology. Background Technology

[0002] In JPEG compression, given a color image quantized in a compressed color space, JPEG image compression color space estimation involves estimating an unknown compressed color space from that image. Suppose the image initially resides in an unknown color space PQR, and after JPEG compression, it enters an arbitrary color space ABC. We need to estimate the linear transformation between the PQR and ABC color spaces. This knowledge is crucial for color image enhancement and JPEG recompression.

[0003] This invention uses reinforcement learning to optimize the LLL algorithm in the color space estimation step of image compression. The LLL algorithm is a classic lattice basis reduction algorithm. The LLL algorithm involves the problem of vector ordering. This invention uses reinforcement learning to optimize it. In the original LLL algorithm, when the Lovász condition is not satisfied, two adjacent vectors are simply swapped. However, the optimized algorithm will continue to move the previous vector to a suitable position. Experiments show that this invention can reduce the number of execution rounds of the algorithm, improve the running efficiency of the algorithm, and improve the accuracy of the algorithm. Summary of the Invention

[0004] To address the aforementioned problems, the present invention aims to provide a JPEG image compression color space estimation method based on lattice basis reduction. Compared with the traditional LLL algorithm, the LLL algorithm optimized by reinforcement learning can effectively improve the algorithm's running efficiency and accuracy.

[0005] To achieve the above objectives, this invention provides a JPEG image compression color space estimation method based on lattice basis reduction, comprising the following steps:

[0006] S1. Select a color image and perform DCT transformation;

[0007] S2. Select a DCT frequency and obtain a 3D histogram of DCT coefficient vectors based on different 8x8 blocks. Each pixel corresponds to a 3D vector to represent a pixel of any color.

[0008] S3. Sort the histograms obtained in step S2 in descending order;

[0009] S4. Select the item in the sort list. The first vector outside the range is used as the basis vector of the lattice;

[0010] S5. Select the next vector. If there are no remaining vectors in the list, exit.

[0011] S6. Calculate the error between the currently selected vector and the closest vector on the lattice formed by the current basis vectors, and solve it using the lattice basis reduction algorithm;

[0012] S7. Evaluate the error vector calculated in step S6. If the calculated error vector is located in... If the selected vector is outside the range, it is not located in the spanned space of the current basis vector set. The selected vector is then added to the basis vector list, and the lattice basis reduction algorithm is performed on this set of basis vectors, proceeding to step S5. If the calculated error vector is located within... Within the specified range, proceed to step S5;

[0013] S8. Obtain two basis vectors estimated from different DCT frequencies. and ,in It is a color transformation matrix, for and Perform addition and subtraction operations on the columns to eliminate and right and The influence of this is used to obtain the color transformation matrix.

[0014] Furthermore, step S2 involves Discrete Cosine Transform (DCT), which uses 64 basic cosine waves to form an 8*8 block. By giving a value that each basic cosine wave must be multiplied by, the amount of basic cosine wave used is represented. The original 8*8 matrix is ​​transformed into a matrix representing the coefficients corresponding to each cosine wave. These coefficients are the DCT frequencies.

[0015] Furthermore, the error required in step S6 is the nearest vector problem CVP, specifically including:

[0016] S61. Given a set of basis vectors and a target vector Find a linear combination of integers , making The vector formed by Distance to target vector Recently, among them , , It is an m-dimensional target vector. For integer finite fields;

[0017] S62. Using the reinforcement learning-optimized LLL algorithm, output the input basis vector A as an LLL reduced basis vector. ;

[0018] S63. Use the Babai algorithm to process the input... and target vector Output approximate vector ;

[0019] S64. Find the error vector .

[0020] Furthermore, the LLL reduction base in step S62 It needs to satisfy two properties, (1) for any ,have in (2) For any under the Lovász condition, are the coefficients in the Schmidt orthogonalization; have in The Schmitt orthogonalized vector.

[0021] Furthermore, the LLL algorithm in step S62 treats the matrix A as n m-dimensional vectors ( By using Schmidt orthogonalization and a commutation step, an arbitrary basis can be transformed into an LLL reduced basis.

[0022] Furthermore, step S62 specifically includes:

[0023] S621. For the k-th vector Perform Schmidt orthogonalization;

[0024] S622. For the (k-1)th and kth vectors Verify the second property of the LLL reduction basis, namely ;

[0025] S623. If property (2) is satisfied, then continue to check whether the following vectors are satisfied, that is, return to step S621;

[0026] S624. Vector Adjust forward sequentially until a position that satisfies property (2) is reached;

[0027] S625. Based on the previously correctly sorted vectors, use reinforcement learning to determine the vectors. Where should it be moved backward?

[0028] S626. In reinforcement learning, the agent that performs actions or makes decisions is called an agent. The state is a summary of the current environment, the action is the decision made by the agent based on the current state, and the reward is a numerical value returned to the agent by the environment after the agent performs an action. The main learning process of reinforcement learning is to continuously adjust and train the agent based on feedback from the environment. Through continuous interaction and trial and error with the environment, a specific goal is ultimately achieved or the overall efficiency of the action is maximized. The state is set as the difference of the squared lengths of two vectors, the action is set as the number of times the vector is moved backward, and the reward is set based on the correctly ordered vectors.

[0029] S627. Based on the correctly sorted vectors, when the state is s, the corresponding action is a. Then, the reward for taking action a when the action is s is increased by 1. If there are k vectors with state s and action a, then the reward for taking action a when the action is s is increased by k.

[0030] S628. The action is a random action. If the reward value corresponding to all actions is 0, then the action is also randomly selected. Otherwise, the action corresponding to the maximum reward value is selected. If the reward value of multiple actions is the maximum reward value, then the average of these actions is taken.

[0031] S629. When vector After the position is determined, the k-th vector of the new sequence is orthogonalized by Schmitt, that is, return to step S621 and continue the loop until the entire sequence satisfies the property of LLL reduction basis (2).

[0032] Furthermore, in step S63, the approximate vector It is the lattice vector closest to the target vector b.

[0033] Furthermore, the lattice basis reduction algorithm performed on the basis vectors in step S7 is still the LLL algorithm optimized by reinforcement learning.

[0034] Compared with existing technologies, the beneficial effects of this invention are as follows: it can reduce the number of execution rounds of the algorithm, improve the running efficiency and accuracy of the algorithm, and the optimization of the LLL algorithm by reinforcement learning can be extended to other lattice basis reduction algorithms, providing new ideas for solving the LWE problem. Attached Figure Description

[0035] Figure 1 This is a flowchart of a JPEG image compression color space estimation method based on lattice basis reduction.

[0036] Figure 2 This is a flowchart of the LLL algorithm optimized using reinforcement learning in S52.

[0037] Figure 3 This is a comparison diagram between the reinforcement learning optimized algorithm of this invention and the LLL algorithm. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical strategies of the embodiments of the present invention will be described in detail and comprehensively below with the aid of the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments that can be derived by those skilled in the art based on the description of the embodiments of the present invention should be considered to be covered by the protection scope of the present invention.

[0039] like Figures 1 to 2 As shown, this invention proposes a JPEG image compression color space estimation method based on lattice basis reduction. By optimizing the swapping step in the LLL algorithm through reinforcement learning, the number of execution rounds of the original LLL algorithm is effectively reduced, improving the efficiency and accuracy of solving the JPEG image compression color space estimation problem using the LLL algorithm. It has good practical value and mainly includes the following steps:

[0040] S1. Select a color image and perform DCT transformation;

[0041] S2. Select a DCT frequency and obtain a histogram of 3D DCT coefficient vectors based on different 8x8 blocks. Each pixel corresponds to a 3D vector. Since there are 3 color spaces, any color can be formed by adding coefficients to these color spaces. Therefore, a 3D vector can represent a pixel of any color.

[0042] S3. Sort the histograms obtained in step S2 in descending order.

[0043] S4. Select the item in the sort list. The first vector outside the range is used as a basis vector of the lattice. Any vector within the range could be a noisy representation of the origin and is therefore ignored.

[0044] S5. Select the next vector. If there are no remaining vectors in the list, exit.

[0045] S6. Calculate the error between the currently selected vector and the closest vector on the lattice formed by the current basis vectors, and solve it using the lattice basis reduction algorithm.

[0046] S7. Evaluate the error vector calculated in step S6. If the calculated error vector is located in... If the selected vector is outside the range, then it does not lie in the spanned space of the current basis vector set. Therefore, the selected vector is added to the basis vector list, and the lattice basis reduction algorithm is performed on this set of basis vectors, proceeding to step S5; if the error vector lies within... Within the specified range, proceed to step S5;

[0047] S8. Obtain two basis vectors estimated from different DCT frequencies. and ,in It is a color transformation matrix, because the color transformation matrix It is defined as having unit column vectors, so any scaled version such as Both of these problems can be solved. and Perform addition and subtraction operations on the columns to eliminate and right and The influence of this is used to obtain the color transformation matrix.

[0048] The following will provide a detailed explanation of steps S2-S6.

[0049] Step S2 involves Discrete Cosine Transform (DCT), which combines 64 basic cosine waves into an 8x8 block. By providing a value that each basic cosine wave must be multiplied by, the amount of that basic cosine wave used is indicated. The original 8x8 matrix is ​​transformed into a matrix representing the coefficients corresponding to each cosine wave. These coefficients are the DCT frequencies.

[0050] In step S3, the histograms obtained in step S2 are sorted in descending order, which ensures that the lattice basis reduction algorithm starts with the minimum noise vector.

[0051] In step S4, if all elements in the vector are in the range [-5,5], then the vector is regarded as a noisy realization of the origin. Therefore, only vectors outside this range will be added to the basis vectors.

[0052] In step S5, if there are no remaining vectors in the sorting list, the final step of the invention is to deduce the color transformation matrix based on the basis vector sets estimated by different DCT frequencies.

[0053] In step S6, the error is calculated, which is equivalent to solving a nearest vector problem (CVP), specifically including:

[0054] The S61.CVP problem is described as follows: Given a set of basis vectors... and a target vector Find a linear combination of integers , making The vector formed Distance to target vector Recently, among them , , It is an m-dimensional target vector. It is a finite field of integers.

[0055] S62. Using the reinforcement learning-optimized LLL algorithm, output the input basis vector A as an LLL reduced basis vector. .

[0056] S63. Use the Babai algorithm to process the input... and target vector Output approximate vector .

[0057] S64. Find the error vector .

[0058] In step S62, the main purpose of lattice basis reduction is to transform an arbitrary given basis into a set of bases with good orthogonality, and to make each vector as short as possible. The LLL reduction basis needs to satisfy two properties: (1) For any LLL reduction basis, the LLL reduction basis must satisfy two properties: (2) For any LLL reduction basis, the LLL reduction basis must satisfy two properties: (3) For any LLL reduction basis, the LLL reduction basis must satisfy two properties: (4) For any LLL reduction basis, ,have in (2) For any under the Lovász condition, are the coefficients in the Schmidt orthogonalization; have in The Schmitt orthogonalized vectors. Specifically, the LLL algorithm treats matrix A as n m-dimensional vectors ( By using Schmidt orthogonalization and a commutation step, an arbitrary basis can be transformed into an LLL reduced basis.

[0059] Step S62 specifically includes:

[0060] S621. For the k-th vector Perform Schmidt orthogonalization;

[0061] S622. For the (k-1)th and kth vectors Verify the second property of the LLL reduction basis, namely ;

[0062] S623. If property (2) is satisfied, then continue to check whether the following vectors are satisfied, that is, return to step S621;

[0063] S624. Vector Adjust forward sequentially until a position that satisfies property (2) is reached;

[0064] S625. Based on the previously correctly sorted vectors, use reinforcement learning to determine the vectors. Where should it be moved backward?

[0065] S626. In reinforcement learning, the agent that performs actions or makes decisions is called an agent. The state is a summary of the current environment, the action is the decision made by the agent based on the current state, and the reward is a numerical value returned to the agent by the environment after the agent performs an action. The main learning process of reinforcement learning is to continuously adjust and train the agent based on feedback from the environment. Through continuous interaction and trial and error with the environment, it ultimately achieves a specific goal or maximizes the overall efficiency of the action. The state is set as the difference of the squared lengths of two vectors, the action is set as the number of times the vector is moved backward, and the reward is set based on the correctly ordered vectors.

[0066] S627. Based on the correctly sorted vectors, when the state is s and the corresponding action is a, the reward for taking action a when the action is s is increased by 1. If there are k vectors with state s and action a, the reward for taking action a when the action is s is increased by k. For example, if there are pairwise combinations of 3 vectors ( If the squared difference of the length of a vector is 1 (state 1), then the reward for state 1 and action of swapping 0 vectors backward is increased by 3.

[0067] S628. When selecting an action, there is a certain probability of randomly selecting an action to ensure that the agent has the opportunity to discover new exchange methods. If the reward value corresponding to all actions is 0, then the action is also randomly selected; otherwise, the action corresponding to the maximum reward value is selected. If multiple actions have the maximum reward value, then the average of these actions is taken.

[0068] S629. Vector After the position is determined, the k-th vector of the new sequence is orthogonalized by Schmitt, that is, return to step S621 and continue the loop until the entire sequence satisfies the property of LLL reduction basis (2).

[0069] In step S63, the Babai algorithm is a lattice-based approximation algorithm that can be used to solve the approximation-CVP problem. Its basic idea is to project the input vector onto the basis vectors of the lattice, and then round the projected vector to obtain an approximate nearest vector. The input is the basis vectors and the target vector, and the output is the lattice vector closest to the target vector. Specifically, the input... and target vector The output is an approximate vector. .

[0070] The pseudocode is as follows:

[0071] In step S7, if the error vector calculated in step S6 is located in If the vector is outside the range, then the currently selected vector does not lie in the spanned space of the current basis vector set. The vector space formed by all linear combinations of vectors is called the spanned space, and the spanned space of basis vectors is the lattice corresponding to the basis vector set. The set of all linear combinations of a set of vectors with integer coefficients is called the lattice L, and the linearly independent set of vectors constituting the lattice is called a basis of the lattice L; if the error vector lies in... If the selected vector is within the range, then the currently selected vector lies in the spanned space of the current basis vector set, and step S5 is executed.

[0072] In step S8, by... and Perform addition and subtraction operations on the columns to eliminate and right and The impact, received and Then, normalize these two matrices and take their average. Normalization involves dividing each element of the matrix by the sum of all its elements and converting the column vectors to unit column vectors, such that the sum of the squares of all elements in each column equals 1. In practice, the column order and sign of the matrix can be changed.

[0073] Explanation of relevant terms:

[0074] Schmidt orthogonalization: a method to transform a set of linearly independent vectors into a set of orthogonal vectors, the process of which is as follows:

[0075] ,

[0076] ,

[0077] ,

[0078] ,

[0079] ...

[0080]

[0081] in, .

[0082] The following will illustrate this with specific examples.

[0083] The practical problem discussed here is set as follows: assuming a color space conversion from ITU.BT-601 YCbCr to RGB, what is the actual color conversion matrix? The transformation matrix needs to be obtained through the following steps. The specific steps are as follows:

[0084] S1. Select a color image and perform DCT transformation;

[0085] S2. Divide the JPEG compressed image into several 8*8 blocks. Each pixel in each block can be represented as a three-dimensional vector. Obtain a three-dimensional histogram based on these vectors.

[0086] S3. Sort the histogram obtained in step S2 in descending order to facilitate the selection of vectors in the following steps;

[0087] S4. Select the item in the sort list. The first vector outside the range is used as a basis vector of the lattice. Any vector within the range could be a noisy representation of the origin, and therefore is ignored;

[0088] S5. Select the next vector in the order arranged in step S3. If there are no remaining vectors in the list, exit.

[0089] S6. Calculate the error between the currently selected vector and the closest vector on the lattice formed by the current basis vectors, using the reinforcement learning-optimized LLL algorithm and Babai algorithm.

[0090] First, the reinforcement learning-optimized LLL algorithm is used to output the input basis vector A as an LLL reduced basis vector. Then, using the Babai algorithm, the input... and target vector Output approximate vector Finally, calculate the error vector. .

[0091] Since the initial matrix A is a poor basis in lattice basis reduction, it needs to be transformed into a good basis, that is, a basis that is both orthogonal and short. The LLL algorithm is such a lattice basis reduction algorithm, and this invention uses reinforcement learning to optimize it.

[0092] Initially, k is set to 2, for the second vector Perform Schmidt orthogonalization. Then... Verify the Lovász condition; if satisfied, then k+1, for the third vector. Perform Schmidt orthogonalization and continue with... Verify the Lovász condition; if it is not satisfied, then... Move to At this point, there is only one sequence that satisfies the condition. Since the rewards in reinforcement learning cannot be updated, random selection is used. Move back a few places, then match with exchange Perform Schmidt orthogonalization and continue with... , Verify the Lovász condition, at this time It is the first vector. It is the second vector.

[0093] When there are two or more vectors in the preceding sequence that satisfy the Lovász condition, the reward can be updated based on that sequence. When the action corresponding to state s is 'a', the reward is incremented by 1; if state s has k corresponding actions of 'a', the reward is incremented by k. For example, when... and If the condition is not met, then the preceding sequence is considered. To update the reward, if the pairwise combinations of the three vectors ( If the squared difference of the lengths of ) is 1 (state 1), then the reward for state 1 and action of swapping 0 vectors backward is increased by 3; if there are pairwise combinations of 4 vectors ( If the squared difference of the length of state l is l, then the reward for state l and action of swapping one vector backward is increased by 4.

[0094] When selecting actions, there is a certain probability of random selection to ensure the agent has a chance to discover new swapping methods. If the reward value for all actions is 0, then an action is also randomly selected; otherwise, the action with the highest reward value is selected. If multiple actions have the highest reward value, then the average of these actions is taken. For example, if swapping 2, 5, or 10 vectors backwards all correspond to the highest reward value, then the number of vectors swapped backwards is... , [] represents rounding, which means changing the vector by 6.

[0095] When the loop reaches the end of the sequence, it means that the entire vector sequence satisfies the Lovász condition, and at this point, any arbitrary basis is transformed into an LLL reducing basis.

[0096] like Figure 3 As shown, comparing the method used in this invention with the original LLL algorithm, it can be seen that the number of execution rounds of the algorithm is significantly reduced.

[0097] At this point, a good basis has been obtained through the lattice basis reduction algorithm. The obtained A_LLL reduced basis and the target vector b are input into the Babai algorithm. The vector b is projected onto each basis vector represented by the LLL reduced basis, and the projected vector is rounded to obtain an approximate lattice vector b'. The specific execution steps are based on the following pseudocode.

[0098] S7. If the error vector calculated in step S6 is located in If the selected vector is outside the range, it does not lie in the spanned space of the current basis vector set. Therefore, the selected vector is added to the basis vector list, and the reinforcement learning optimization LLL algorithm is performed on this set of basis vectors. Execute step S5; if the error vector lies within... If the selected vector is within the range, then the currently selected vector lies in the spanned space of the current basis vector set, and step S5 is executed.

[0099] S8. After the above steps, we obtain two basis vector sets, which can be represented in matrix form as follows: and While the first two columns of these two matrices are scaled versions of each other, the third column is not. This problem can be easily solved by adding the first column to the corresponding third column. At this point, the two matrices become... and The color conversion matrix is ​​obtained by normalizing the above matrix and averaging it. The estimated transformation matrix can be seen. Compared with the original color transformation matrix Almost identical, Columns 2 and 3 were swapped, and the symbols were reversed for easier identification. Comparison. In practice, the changes in ordering and sign are negligible.

[0100] Finally, it should be specifically pointed out that the above embodiments are only used to illustrate the technical solution of the present invention and are not intended to limit it. Various adjustments and improvements made by those skilled in the art to the technical solution of the present invention without departing from the design concept of the present invention should be included within the protection scope of the present invention.

Claims

1. A JPEG image compression color space estimation method based on lattice basis reduction, characterized in that, Includes the following steps: S1. Select a color image and perform DCT transformation; S2. Select a DCT frequency and obtain a 3D histogram of DCT coefficient vectors based on different 8x8 blocks. Each pixel corresponds to a 3D vector to represent a pixel of any color. S3. Sort the histograms obtained in step S2 in descending order; S4. Select the item in the sort list. The first vector outside the range is used as the basis vector of the lattice; S5. Select the next vector. If there are no remaining vectors in the list, exit. S6. Calculate the error between the currently selected vector and the closest vector on the lattice formed by the current basis vectors, using the lattice basis reduction algorithm; specifically including: S61. Given a set of basis vectors and a target vector Find a linear combination of integers , making The vector formed by Distance to target vector Recently, among them , , It is an m-dimensional target vector. For integer finite fields; S62. Using the reinforcement learning-optimized LLL algorithm, output the input basis vector A as an LLL reduced basis vector. The LLL reduction base It needs to satisfy two properties: (1) For any ,have in (2) For any under the Lovász condition, are the coefficients in the Schmidt orthogonalization; have in The Schmitt orthogonalization vector; step S62 specifically includes: S621. For the k-th vector Perform Schmidt orthogonalization; S622. For the (k-1)th and kth vectors Verify the second property of the LLL reduction basis, namely ; S623. If property (2) is satisfied, then continue to check whether the following vectors are satisfied, that is, return to step S621; S624. Vector Adjust forward sequentially until a position that satisfies property (2) is reached; S625. Based on the previously correctly sorted vectors, use reinforcement learning to determine the vectors. Where should it be moved backward? S626. In reinforcement learning, the subject that makes actions or decisions is called an agent. The state is a summary of the current environment, the action is a decision made by the agent based on the current state, and the reward is a value returned to the agent by the environment after the agent performs an action. The state is set as the difference of the squared lengths between two vectors, the action is set as the number of vectors to be swapped, and the reward is set according to the previously correctly sorted vectors. S627. Based on the correctly sorted vectors, when the state is s, the corresponding action is a. Then, the reward for taking action a when the action is s is increased by 1. If there are k vectors with state s and action a, then the reward for taking action a when the action is s is increased by k. S628. The action is a random action. If the reward value corresponding to all actions is 0, then the action is also randomly selected. Otherwise, the action corresponding to the maximum reward value is selected. If the reward value of multiple actions is the maximum reward value, then the average of these actions is taken. S629. When vector After the position is determined, the kth vector of the new sequence is orthogonalized by Schmitt, that is, return to step S621 and continue the loop until the entire sequence satisfies the property of LLL reduction basis (2). S63. Use the Babai algorithm to process the input... and target vector Output approximate vector ; S64. Find the error vector ; S7. Evaluate the error vector calculated in step S6. If the calculated error vector is located in... If the selected vector is outside the range, it is not located in the spanned space of the current basis vector set. The selected vector is then added to the basis vector list, and the lattice basis reduction algorithm is performed on this set of basis vectors, proceeding to step S5. If the calculated error vector is located within... Within the specified range, proceed to step S5; S8. Obtain two basis vectors estimated from different DCT frequencies. and ,in It is a color transformation matrix, for and Perform addition and subtraction operations on the columns to eliminate and right and The influence of this is used to obtain the color transformation matrix.

2. The JPEG image compression color space estimation method according to claim 1, characterized in that: Step S2 involves Discrete Cosine Transform (DCT), which combines 64 basic cosine waves into an 8*8 block. By giving a value that each basic cosine wave must be multiplied by, the amount of basic cosine wave used is represented. The original 8*8 matrix is ​​transformed into a matrix representing the coefficients corresponding to each cosine wave. These coefficients are the DCT frequencies.

3. The JPEG image compression color space estimation method according to claim 1, characterized in that: The LLL algorithm in step S62 treats the matrix A as n m-dimensional vectors ( By using Schmidt orthogonalization and a commutation step, an arbitrary basis can be transformed into an LLL reduced basis.

4. The JPEG image compression color space estimation method according to claim 1, characterized in that: In step S63, the approximate vector It is the lattice vector closest to the target vector b.

5. The JPEG image compression color space estimation method according to claim 1, characterized in that: The lattice basis reduction algorithm performed on the basis vectors in step S7 is still the LLL algorithm optimized by reinforcement learning.

Images