CN109035178B - Multi-parameter value tuning method applied to image denoising - Google Patents

Abstract

The invention discloses a multi-parameter value tuning method applied to image denoising, which gives a target complexity C_budgetUnder the condition of keeping the values of other parameters unchanged, respectively calculating the energy efficiency ratio of the current value of a plurality of parameters to the next candidate value of the parameters,determining the parameter with the maximum energy efficiency ratio, and adjusting the candidate value of the parameter to be the current value, namely completing the parameter tuning once; and (3) calculating the current parameter combination complexity every time parameter tuning is finished, judging whether the current parameter combination complexity meets the target complexity constraint, if not, executing the steps to continue parameter tuning until the given complexity constraint is met, and stopping tuning.

Description

Multi-parameter value tuning method applied to image denoising

Technical Field

The invention belongs to the field of image processing, and particularly relates to a multi-parameter value tuning method applied to image denoising.

Background

In the field of image processing, image denoising has been a major research point and difficulty. It is a precondition for performing subsequent processing such as image segmentation and feature extraction. At present, most image denoising algorithms continuously pursue denoising performance, and the consideration of algorithm calculation complexity is omitted. In practical applications, the complexity of the algorithm is also an important factor. Different application scenarios have different requirements on algorithm performance and complexity. For example, when we want image denoising to be real-time, the high complexity algorithm cannot meet our needs. Therefore, we need to optimize the denoising algorithm with higher complexity.

An algorithm framework based on NLM (non-local means) and iteration is generally applied to an image denoising algorithm, and the performance is better but the complexity is higher. Non-local mean (NLM) algorithms exploit the information redundancy of natural images. When the image is denoised, the noise in the image can be better removed by utilizing the pixels of the neighborhood positions around the pixel points and considering other positions of the non-neighborhood simultaneously. And the iteration can solve the over-smooth phenomenon caused by image denoising, and prevent the image from losing edge and texture information. And carrying out weighted summation on the denoised image and the noise image to obtain a new noise-containing image, and starting the next NLM denoising iteration. And obtaining a final denoised image after multiple iterations. This iterative approach can eliminate image noise while preserving image detail well.

The final complexity and denoising performance expression of the frame denoising algorithm are determined by the common influence of a series of algorithm key parameters. Different value combinations of the parameters are realized, and the algorithm shows different computational complexity and denoising performance. In practical applications, the computational complexity of the denoising algorithm is usually limited. Given the computational complexity of a demand application scenario, it is a common problem how to quickly and accurately determine the values of a plurality of parameters of an algorithm. An exhaustive search method can be adopted to try all possible candidate values of multiple parameters to find a parameter combination with better performance and lower complexity, but the search process is time-consuming and labor-consuming, and the result has no generality. These parameters are currently configured essentially empirically in the literature.

Disclosure of Invention

Aiming at the problem, the invention provides a multi-parameter value tuning method applied to image denoising, which can quickly and effectively optimize parameters and find a parameter combination with better performance and lower complexity. Typical multi-parameter image denoising algorithms include NCSR, SSC-GSM, BM3D algorithms and the like.

The invention provides a multi-parameter value tuning method applied to image denoising, which gives a target complexity C_budgetUnder the condition that the values of other parameters are kept unchanged, respectively calculating the energy efficiency ratios of the current values of the parameters and the next candidate value of the parameters, determining the parameter with the maximum energy efficiency ratio, and adjusting the candidate value to the current value, namely completing one-time parameter tuning;

and (3) calculating the current parameter combination complexity every time parameter tuning is finished, judging whether the current parameter combination complexity meets the target complexity constraint, if not, executing the steps to continue parameter tuning until the given complexity constraint is met, and stopping tuning.

Further, the method specifically comprises the following steps:

a) initializing parameters: suppose there are M parameters P to be optimized₁,P₂,…,P_m,…,P_MThe mth parameter has n_mA number of possible discrete candidate values; k is a radical of_mIndex value, k, for the value of the mth parameter_m＝1～n_mA certain value; the value of each parameter is recorded as V_km(ii) a Setting each parameter as a default value of the algorithm, and taking the default value as an initial value of parameter tuning; and calculating the performance q and the complexity C of the default value combination. Given target complexity C_budget；

b) Calculating current parametersEnergy efficiency ratio of (d); for the current parameter combination, its performance q (k) is calculated_m) And complexity C (k)_m) (ii) a Under the condition of keeping the values of other parameters unchanged, changing the value of one parameter every time, changing the current value into the next candidate value, and calculating the performance q (k'_m) And complexity C (k'_m) (ii) a Then respectively calculating the current value of each parameter and the next discrete candidate value k'_mEnergy efficiency ratio τ of

Δq(k_m)＝q(k_m)-q(k′_m) (1)

ΔC(k_m)＝C(k_m)-C(k′_m) (2)

τ(k_m)＝Δq(k_m)/ΔC(k_m) (3)

c) For each parameter τ (k)_m) Sorting in a descending order; τ (k)₁),τ(k₂),…,τ(k_M) After sorting, the index is changed from the original (1,2, …, M) to (l)₁,l₂,…，l_M) Here (l)₁,l₂,…，l_M) Is some new ordering of (1,2, …, M), i.e. /)₁Indexing the parameter with the maximum energy efficiency ratio; let i be the index of the current optimized parameter, i ═ l₁Performing the ith parameter value optimization P_i；

d) Calculating the algorithm complexity C 'and the performance q' of a parameter combination when the ith parameter takes a candidate value;

e) judging whether the calculation complexity C' is larger than the target complexity C_budget(ii) a If the condition is not met, the current adjustment is effective, and the adjustment is further carried out according to the trend to change the P of the ith parameter_iIs changed to its candidate value, V_ki＝V_k'_iI.e. k_i＝k'_iEntering step b) to continue adjustment; if the condition is met, ending the parameter value optimizing process and outputting the current parameter values.

The multi-parameter value tuning method applied to image denoising adjusts multi-control parameter values for a multi-parameter image denoising algorithm with high complexity and high performance, and reduces algorithm complexity. According to the method, firstly, a target computation complexity constraint is given, and then the multi-parameter value is adjusted through an adjusting and optimizing method according to the given complexity constraint, so that the optimal denoising performance under the computation complexity constraint is obtained. The invention is inspired by the gradient descent thought in deep learning, and simply and effectively realizes the multi-parameter value optimization of the algorithm by calculating the ratio of the complexity of different values of the same parameter to the performance change, namely the efficiency ratio, as a judgment basis when the parameter is adjusted and optimized.

Drawings

FIG. 1 is a schematic diagram of a multi-parameter value combination;

FIG. 2 is a schematic diagram of constrained optimization parameter combination selection;

fig. 3 is a flow chart of a multi-parameter tuning method.

FIG. 4 is a key parameter diagram;

fig. 5 a-5 d are graphs of performance versus complexity for various key parameters.

Detailed Description

The invention will be further explained with reference to the drawings.

As shown in FIG. 1, for the image denoising algorithm commonly affected by a plurality of key parameters, it is assumed that there are M parameters P to be optimized₁,P₂,…,P_m,…,P_MEach having n_mAnd taking the value of each candidate parameter. The m-th parameter having n_mA possible candidate value (discrete value). k is a radical of_mIndex value (k) for the value of the mth parameter_m＝1～n_mSome value in). The value of each parameter is recorded as V_km。

Given target complexity C_budgetEach parameter P needs to be given_mAnd selecting proper values to ensure that the average denoising performance (a plurality of training images) of the multi-parameter value combination is optimal, namely PSNR or SSIM.

The multi-parameter value tuning method applied to image denoising, provided by the invention, has the advantages that the target complexity constraint of the algorithm is given, the multi-parameter value is optimized under the constraint, and the parameter combination with the optimal performance is found.

Given target complexity C_budgetUnder the condition that the values of other parameters are kept unchanged, the energy efficiency ratios of the current values of the parameters and the next candidate value of the parameters are respectively calculated, the parameter with the largest energy efficiency ratio is determined, the candidate value of the parameter is adjusted to be the current value, and then one-time parameter tuning is completed.

And (3) calculating the current parameter combination complexity every time parameter tuning is finished, judging whether the current parameter combination complexity meets the target complexity constraint, if not, executing the steps to continue parameter tuning until the given complexity constraint is met, and stopping tuning.

Examples

The existing multi-parameter image denoising algorithm has high complexity and cannot be applied to a real-time or short-time-delay scene. The method and the device reduce the algorithm complexity by optimizing multi-parameter values based on target complexity constraint. As shown in fig. 3, it is a specific analysis and tuning process of a typical multi-parameter image denoising algorithm.

Algorithmic parameter analysis

The image denoising algorithm has a plurality of parameters, the complexity or performance of the algorithm is only affected by different values of some parameters, and the parameters can be obtained by taking values with lower complexity or higher performance. The invention mainly analyzes parameters which simultaneously affect the complexity and the performance of the algorithm. The different discrete values of the parameters enable the complexity and performance of the algorithm to form positive correlation, such as the size N of the image block, the search range when the similar block is searched by stepping step and NLM, and the number N of the similar blocks_blkThe iteration times of the iterative denoising algorithm and the like. A schematic of these key parameters is shown in fig. 4.

Fig. 5 shows the relationship between the performance and complexity of each key parameter of an algorithm. Within a certain range, different values of key parameters in the algorithm often make complexity and performance in a positive correlation relationship. The process of optimal selection of algorithm parameters is therefore actually a trade-off between performance q and complexity C. When the parameter value changes, the change of the complexity C and the change of the performance q are caused, and the significance degree of the value change of different parameters of the same algorithm on the q and the C is different. Therefore, the ratio tau of the change of q and C, namely the efficiency ratio, is used as the criterion of the performance when the parameters are adjusted to be optimal.

Specific implementation process of algorithm

a) Initializing parameters; suppose there are M parameters P to be optimized₁,P₂,…,P_m,…,P_MThe mth parameter has n_mA number of possible discrete candidate values. k is a radical of_mIndex value (k) for the value of the mth parameter_m＝1～n_mSome value in). The value of each parameter is recorded as V_km. And setting each parameter as a default value of the algorithm to serve as an initial value of parameter tuning. And calculating the performance q and the complexity C of the default value combination. Given target complexity C_budget；

b) Calculating the energy efficiency ratio of each current parameter; for the current parameter combination, its performance q (k) is calculated_m) And complexity C (k)_m) (ii) a Under the condition of keeping the values of other parameters unchanged, changing the value of one parameter every time, changing the current value into the next candidate value, and calculating the performance q (k'_m) And complexity C (k'_m) (ii) a Then respectively calculating the current value of each parameter and the next discrete candidate value k'_mEnergy efficiency ratio τ of

Δq(k_m)＝q(k_m)-q(k′_m) (1)

ΔC(k_m)＝C(k_m)-C(k′_m) (2)

τ(k_m)＝Δq(k_m)/ΔC(k_m) (3)

c) For each parameter τ (k)_m) Sorting in a descending order; τ (k)₁),τ(k₂),…,τ(k_M) After sorting, the index is changed from the original (1,2, …, M) to (l)₁,l₂,…，l_M) Here (l)₁,l₂,…，l_M) Is some new ordering of (1,2, …, M), i.e. /)₁And indexing the parameter with the largest energy efficiency ratio. Let i be the index of the current optimized parameter, i ═ l₁Performing the ith parameter value optimization P_i. Because of τ (k)_i) The maximum, unit computational complexity performance gain is greatest.

d) Calculating the algorithm complexity C 'and the performance q' of a parameter combination when the ith parameter takes a candidate value;

e) judging whether the calculation complexity C' is larger than the target complexity C_budget(ii) a If the condition is not met, the current adjustment is effective, and the adjustment is further carried out according to the trend to change the P of the ith parameter_iIs changed to its candidate value, V_ki＝V_k'_i(i.e. k)_i＝k'_i) Entering the step b to continue adjustment; if the condition is met, ending the parameter value optimizing process and outputting the current parameter values.

Technical effects

(1) In the process of parameter optimization and adjustment, whether the algorithm complexity realized by multi-parameter combination at the moment accords with the target complexity constraint or not is judged every time parameter optimization is completed, so that the parameter combination complexity obtained based on the method accords with the target complexity constraint condition.

(2) Because the default values of the image denoising algorithm are all parameter combinations with high complexity and high performance, the default values are set as the initial values of the algorithm, and the optimization is carried out according to the method of the invention, which is a process for reducing the complexity and the performance. Therefore, the performance of the algorithm is optimal under the constraint condition of meeting the given complexity;

as shown in fig. 2, the final implementation effect is that, under the same complexity constraint, the performance of the algorithm implemented by the parameter value combination of the multi-parameter value tuning method is optimal.

Claims (1)

1. A multi-parameter value tuning method applied to image denoising is characterized by comprising the following steps: given target complexity C_budgetUnder the condition that the values of other parameters are kept unchanged, respectively calculating the energy efficiency ratios of the current values of the parameters and the next candidate value of the parameters, determining the parameter with the maximum energy efficiency ratio, and adjusting the candidate value to the current value, namely completing one-time parameter tuning;

when parameter tuning is finished once, calculating the complexity of the current parameter combination, judging whether the complexity of the parameter combination at the moment meets the target complexity constraint, if not, executing the steps to carry out parameter tuning continuously until the given complexity constraint is met, and stopping adjusting;

the method specifically comprises the following steps:

a) initializing parameters: suppose there are M parameters P to be optimized₁,P₂,…,P_m,…,P_MThe mth parameter has n_mA number of possible discrete candidate values; k is a radical of_mIndex value, k, for the value of the mth parameter_m＝1～n_mA certain value; the value of each parameter is recorded as V_km(ii) a Setting each parameter as a default value of the algorithm, and taking the default value as an initial value of parameter tuning; calculating the performance q and the complexity C of default value combination, and giving a target complexity C_budget；

b) Calculating the energy efficiency ratio of each current parameter; for the current parameter combination, its performance q (k) is calculated_m) And complexity C (k)_m) (ii) a Under the condition of keeping the values of other parameters unchanged, changing the value of one parameter every time, changing the current value into the next candidate value, and calculating the performance q (k'_m) And complexity C (k'_m) (ii) a Then respectively calculating the current value of each parameter and the next discrete candidate value k'_mEnergy efficiency ratio τ of

Δq(k_m)＝q(k_m)-q(k′_m) (1)

ΔC(k_m)＝C(k_m)-C(k′_m) (2)

τ(k_m)＝Δq(k_m)/ΔC(k_m) (3)

c) For each parameter τ (k)_m) Sorting in a descending order; τ (k)₁),τ(k₂),…,τ(k_M) After sorting, the index is changed from the original (1,2, …, M) to (l)₁,l₂,…，l_M) Here (l)₁,l₂,…，l_M) Is some new ordering of (1,2, …, M), i.e. /)₁Indexing the parameter with the maximum energy efficiency ratio; let i be the index of the current optimized parameter, i ═ l₁Performing the ith parameter value optimization P_i；

d) Calculating the algorithm complexity C 'and the performance q' of a parameter combination when the ith parameter takes a candidate value;

e) judging whether the calculation complexity C' is larger than the target complexity C_budget(ii) a If the condition is not met, the current adjustment is effective, and the adjustment is further carried out according to the trend to change the P of the ith parameter_iIs changed to its candidate value, V_ki＝V_k'_iI.e. k_i＝k'_iEntering step b) to continue adjustment; if the condition is met, ending the parameter value optimizing process and outputting the current parameter values.

Images