Application of non-local mean image denoising algorithm based on machine learning technology in visual communication design

Abstract

In Visual Communication Design (VCD), noise data is easy to appear, which reduces image quality and affects the effect of VCD. The non local mean image denoising algorithm is a good filtering denoising algorithm, but there are still issues of information interference and missing. To improve the performance of noise recognition and image denoising technology, this study proposes a non local mean image denoising algorithm based on machine learning technology. The whale optimization algorithm, as a machine learning technique, has good performance in seeking optimal solutions. Therefore, it is applied to optimize the filtering parameters of non local mean image denoising algorithms to improve the perforGAmance of non local mean image denoising algorithms. To address the shortcomings of the whale optimization algorithm, BP neural network is introduced for optimization. Finally, the experiment uses the improved particle swarm optimization algorithm to optimize the BPNN and applies it to the recognition and classification of noise data. Combining the above contents, the IBINLM image denoising algorithm is constructed experimentally. It is verified that the IPSO-BPNN model’s loss value is 0.12; The recognition accuracy of the model for noise pixels is 98.64%; F1 value reaches 96.32%; The fitting degree reaches 0.983. The PSNR of IBINLM algorithm is 35.86 dB; MSE is 0.29; AUC value reaches 0.903. The results show that the IPSO-BPNN model and IBINLM image denoising algorithm have better performance compared to other models, which can improve the quality of visual communication works, playing an essential role in image transmission and storage in visual communication design.

Keywords

Machine learning non-local mean image denoising whale optimization algorithm visual communication design

1 Introduction

In the storage and transmission of VCD, noise data is easy to be generated due to various reasons, which reduces the image quality and affect the effect of VCD. Image denoising technique can filter and remove the noise in image data, thus improving the purity and quality of image data [1]. Therefore, image denoising is very important. Currently, common image denoising technologies include median filtering, mean filtering, median filtering and adaptive Wiener filtering. However, the denoising effect and efficiency of these methods need to be improved. A quantity of time is taken to process large-scale image data, so it is gradually difficult to adapt to the needs of VCD [2, 3]. In the development of information technology, to enhance data processing efficiency and realize automation and intelligence, the non local mean image denoising algorithms come into being and developed rapidly in recent decades, becoming a research hotspot. Combining machine learning technology with non local mean image denoising algorithms to improve image denoising performance has been a popular research direction in recent years [4, 5]. Therefore, to improve the denoising performance of non local mean image denoising algorithms based on machine learning technology, this study proposes an optimized non local mean (NLM) image denoising algorithm based on an improved BP neural network (BPNN), which utilizes the characteristics of BP neural network to achieve large-scale image noise recognition, thereby improving the image denoising efficiency and accuracy of NLM. This study improves the denoising performance of non local mean image denoising algorithms, providing ideas and data support for the application of image denoising technology in visual communication.

2 Related works

In the current Internet environment, image is an essential way to transmit information, which can display various kinds of information for people intuitively, efficiently and vividly. However, due to various reasons, most of the image data have different amounts of noise, which has a negative impact on the image’s quality and the acquisition of information. Therefore, image denoising is very essential. With the advancement of deep learning and machine learning technology, more researchers begin to apply deep learning and machine learning technique to the field of image denoising. C Liu et al. proposed a method based on wavelet transform and non local moment mean filtering, and used wavelet based soft threshold denoising technology to denoise noisy images. Experiments showed that this algorithm achieved better denoising performance compared to other denoising methods [6]. In view of the defect that the current image denoising technology can only reduce the noise of a few images at the same time and has low efficiency, Tian et al. combined with the depth convolution neural network (DCNN) to build an image denoising model that could process images in large quantities. This model greatly improved the efficiency of image denoising [7]. Gurrola-Ramos et al. analyzed the use and performance of residual dense u neural network, and studied its application effect in image denoising, providing a new path for the development of image denoising technology [8]. Shi et al. extracted and fused multi-scale features, and introduced attention mechanism in CNN network to avoid defects in existing hyperspectral image denoising technology, to improve the effect of hyperspectral image denoising [9]. Xie et al. proposed solutions and optimization strategies for the deficiencies in the current research results related to the self-supervised denoising framework. The proposed strategy improved the image denoising performance of the self-supervised denoising framework [10]. Valsesia et al. introduced the depth map convolution neural network (DGCN) in detail, analyzed the performance and use of DGCN, and applied DGCN to the image denoising. The experimental results showed that the model had good denoising effect [11]. Zhang et al. constructed a technology that introduced attention-guided scaling mechanism, aiming at the challenging problems widely existing in the current image denoising. In the experiment, the denoising performance of this method was significantly higher than the existing image denoising technology [12]. Huang et al. constructed a reversible network structure based on wavelet heuristic algorithm. The reversible network structure was applied to the field of image denoising. The results indicated that the method had good performance and high practicability [13].

In the development of information technology to improve the efficiency of data processing, machine learning technique came into being. In recent decades, its rapid development has become a hot spot in the research field. Currently, machine learning technology is widely used in various fields and has made great contributions to the development of digitalization and intelligence in all walks of life. Ezugwu A E et al. proposed an optimization algorithm for grassland dogs based on their predatory behavior. This algorithm could better optimize based on the predation and aggregation behavior of grassland dogs. The experimental results showed that compared with existing algorithms, this algorithm could effectively find the global optimal solution [14]. Agushaka J O et al. proposed a dwarf mongoose optimization algorithm, introducing a new operator to enhance the algorithm’s search ability. The performance of IDMO was compared with DMO and eight other existing algorithms using different performance indicators and statistical analysis. In most cases, the results indicated that the solutions obtained by IDMO were superior to those obtained by existing algorithms [15]. Agushaka J O et al. proposed a gazelle optimization algorithm based on the behavior of gazelles during predation. The algorithm simulated the escape behavior of a gazelle when it found a predator and searched for the optimal solution. The simulation results confirmed the superiority and competitiveness of the Gazelle optimization algorithm compared to the nine state-of-the-art algorithms currently available in the literature [16]. Chan et al. reviewed and discussed comprehensively the literature on the application of machine learning technology to analyze big data and customer emotions. This research provided theoretical support for the application and development of big data and machine learning technique [17]. Aggarwal et al. analyzed the current application of machine learning and its role in medical care. The analysis results showed that machine learning technology could improve medical care methods when facing COVID-19 [18]. Stojadinovi and others combined machine learning technology to build a classification model. The experiment applied it to the prediction of the loss in the post-earthquake disaster area, providing data support for the rapid reconstruction of the earthquake area [19]. Ma combined the computing power of machine learning with human insight, thus applying machine learning technology to marketing. This method could enhance the effect of marketing [20]. Meuwly used machine learning technology to quickly analyze and calculate chemical reactions. It was showed that machine learning could enhance the efficiency and accuracy of chemical reaction analysis [21].

In summary, the current research on the integration of various deep learning methods and image denoising techniques is relatively mature and has important applications in various fields. However, there is relatively little research on the fusion of image denoising technology and BP neural network, and there is also a lack of optimization research on the technology. Therefore, this study combines image denoising technology with whale algorithm, optimizing the BP neural network, and proposes a non local mean image denoising algorithm based on machine learning technology, which plays a certain reference role in the field of image denoising in visual communication.

3 Improved NLM algorithm based on machine learning technology

3.1 Optimization of NLM image denoising algorithm based on WOA

Visual communication refers to the efficient and direct visual transmission of the thoughts, feelings and information that designers want to express to the audience through design techniques. Currently, thanks to the Internet technology, VCD has also undergone major changes. Using computer to realize VCD has become the mainstream form. However, in the process of visual communication design storage and transmission, noise data is prone to be generated due to various reasons. Common noise includes brightness noise and color noise generated by image devices. The compression noise generates during the compression process; There is digital noise in image processing. These noises can reduce image quality and affect the effectiveness of visual communication design. Therefore, image denoising plays an essential role in visual communication design [22]. The basic idea of NLM image denoising algorithm is to utilize self similarity in the image. In images, similar structures have similar neighborhood blocks at different positions. The NLM algorithm performs denoising by calculating the similarity between neighboring blocks of each pixel in the image. Specifically, the algorithm first selects a reference block and then searches the entire image for the neighborhood block that is most similar to the reference block. Similarity is measured by calculating the pixel difference between two blocks. The algorithm uses the similarity to determine the weight of the neighborhood block, and then weights the reference block according to the weight of the neighborhood block to obtain the denoising result. By performing similar processing on the pixels of the entire image, the denoising image is ultimately obtained [23]. The principle of NLM is shown in Fig. 1. The similarity of all pixels in the image is calculated through x and y windows, and then different weights are given to each pixel.

Fig. 1

The principle of noise removal of NLM.

If there is an image u (x) without noise, when it is disturbed by noise, Equation (1) express the image f (x) polluted by noise. $f (x) = u (x) + n (x)$ (1)

In Equation (1), n (x) is a noise signal unrelated to u (x). I indicates the image’s pixel gray value; i indicates a pixel point in the image; The filtering form of NLM for noisy images can be expressed by Equation (2). $NL [v] (i) = \frac{1}{C (i)} \sum_{j \in I} w (i, j) v (j)$ (2)

In Equation (2), NL [v] (i) is the value obtained after the weighted average of the target pixel points; C (i) is the normalization constant; v (j) is the gray value of the pixel point j; W (i, j) is the similarity between the neighborhood window of the target pixel point i and the pixel point j, which is calculated by Equation (3). $W (i, j) = \frac{1}{G (i)} exp (- \frac{d (i, j)}{h})$ (3)

In Equation (3), G (i) is the normalization constant; d (i, j) indicates the Gaussian weighted Euclidean distance between two pixel points i, j; h is a filtering parameter, which determines the range of the exponential dependent variable and has a certain impact on the smoothness of the image. Therefore, the selection of filtering parameters often determines the performance of NLM image denoising algorithm. The manual selection of filter parameters is time-consuming and laborious, and the result of parameter selection is often not optimal, which affects the performance of NLM. To ensure the function of NLM algorithm, whale optimization algorithm (WOA) is introduced to select the optimal filtering parameters, so as to optimize NLM algorithm. Figure 2 illustrates the optimization process of WOA

Fig. 2

WOA optimization process.

WOA is suggested in recent years. It mainly simulates the process of whale search and predation to optimize the solution of the target. In the iterative process of WOA, the best prey that whales can prey on is the optimal solution of the optimization issue. In the initial stage, the NLM algorithm is applied to perform initial denoising on the image and obtain an initial solution. WOA algorithm is used as the initial solution of iterative search for individuals in the population. In each iteration, the quality of the solution is evaluated by calculating the fitness function, and the position of the solution is updated based on the WOA search strategy. When updating the position of the solution, local or global search can be selected based on the value of the fitness function. The WOA optimization phase is concluded based on a preset maximum number of iterations [24, 25]. Based on the optimized parameters, the NLM algorithm is applied to perform the final denoising on the image and obtain the optimized results. The mathematical model of WOA is shown in Equations (4) and (5). $D = | C X^{*} (t) - X (t) |$ (4) $X (t + 1) = X^{*} (t) - AD$ (5)

In Equations (4) and (5), D denotes the distance between the whale individual and the optimal individual position under the current iteration number; t is the current iteration; X is the vector representation of whale individual position; X^* is the vector representation of the optimal individual position under the current iteration number; A, C are coefficient vectors, that is, the enclosing step size. In the process of searching and hunting prey, whales emit spiral bubbles to surround prey, which can be expressed as Equations (6) and (7). $D^{'} = | X^{*} (t) - X (t) |$ (6) $X (t + 1) = D^{'} e^{θ l} \cdot cos (2 π l) + X^{*} (t)$ (7)

In Equations (6) and (7), D′ is the distance between the whale individual and the optimal individual position under the current iteration number at the time of iteration; l is a random number with a value range of - 1 to 1; θ is a fixed constant. When searching for prey, the formula is expressed as Equations (8) and (9). $D^{″} = | C \cdot X_{rand} - X (t) |$ (8) $X (t + 1) = X_{rand} - A \cdot D^{″}$ (9)

In Equations (8) and (9), X_rand denotes the position of any random individual in the population; D″ is the distance from X to the corrected random individual position C · X_rand. Through the above operations, the optimal solution obtained is the optimal value of NLM filtering parameters, thus improving the performance of NLM.

3.2 Multi-strategy collaborative improvement WOA

The research uses WOA to optimize NLM to get the best filtering parameters and enhance the image denoising function of NLM. But the WOA algorithm has many defects, such as the accuracy and convergence performance are not ideal, and it is easy to find local optimization. As a result, the filtering parameters obtained through WOA may not be the optimal parameter values, thus affecting the performance of NLM algorithm [26]. Therefore, this study proposes strategies to optimize WOA. Firstly, a population information guidance strategy is proposed. In WOA, the individual with the lowest fitness value in the population is regarded as the best individual, and the global optimal position of the population is mainly generated by the optimal individual generated in the iterative process. In this process, the population is closer and closer to the prey, but the influence of other whale individuals except the optimal one on the global optimal solution is ignored, so it is easy to fall into local behavior. Then, an information guidance strategy is suggested to improve this process, such as Equation (10). $X^{*} (t) = w_{3} X_{1} (t) + w_{2} X_{2} (t) + w_{1} X_{3} (t)$ (10)

In Equation (10), X₁ (t) , X₂ (t) , X₃ (t) represent the optimal position of the population, the suboptimal position of the population and the optimal position of the population at the iteration t; w₁, w₂, w₃ respectively represents the proportion of X₁ (t) , X₂ (t) , X₃ (t) in the t iteration. In Equation (10), the position of a single whale is not matched with its corresponding weight one by one, but is determined according to the inverse ratio of the effect of specific gravity on the solution. Through this strategy, the sharing of location information among different individuals can be strengthened. In addition, the strategy also strengthens the influence of other whale individuals except the optimal one on the global optimal solution, so as to avoid the premature phenomenon of the algorithm. While, this method cannot determine whether the fitness value of the new location is better than the original location. Thus, the greedy rule is introduced to compare the fitness values of the new position and the original position of individuals, and to retain the better individuals. In WOA, when whale individuals surround their prey, they are based on the position and random coefficient of the best individual under the current iteration number, without considering the position of the previous generation of individuals. This causes some individuals in the WOA to be ignored, which affects the accuracy of the WOA. Therefore, WOA is optimized with an improved gold sine idea to enhance its global optimization and convergence performance, as shown in Equation (11).

$\begin{matrix} X (t + 1) = X (t) \cdot | sin (R_{1}) | + R_{2} \\ \cdot sin (R_{1}) \cdot | x_{1} \cdot X^{*} (t) - x_{2} X_{mean} (t) | \end{matrix}$ (11)

In Equation (11), X_mean (t) denotes the average position of the t iteration; R₁ denotes a random number with a value range from 0 to 2π; R₂ is a random number with a value range from 0 to π, which determines the moving distance and direction of the individual whale; The golden section coefficient of x₁, x₂ can guide the whale individual to move to the position of the optimal solution, thus accelerating the convergence of the algorithm. In the iterative process of WOA, the convergence factor a linearly declines as the number of repetitions rises, that is, the global search ability of WOA algorithm decreases linearly. This characteristic will cause the WOA algorithm to be unable to search globally at the beginning of the iteration, and the convergence speed is slow at the end of the iteration. To improve this defect, a nonlinear convergence factor adjustment strategy is proposed, such as Equation (12). $a = {\begin{matrix} \frac{T_{max}^{2}}{4} ({(t - \frac{T_{max}}{2})}^{2} + 1), & t ⩽ \frac{T_{max}}{2} \\ \frac{T_{max}^{2}}{4} {(t - \frac{T_{max}}{2})}^{2}, & t > \frac{T_{max}}{2} \end{matrix}$ (12)

In Equation (12), T_max is the biggest number of iterations. Through Equation (12), the convergence factor of WOA can be adjusted nonlinearly by segments, and the global search ability of WOA early and the convergence performance later can be improved. The change trend of contraction factor between this strategy and the original strategy is shown in Fig. 3.

Fig. 3

Change trend of contraction factor of WOA.

Through the above operations, the optimization of WOA is completed. The optimized WOA is used to optimize NLM, and IWOA-NLM is constructed to improve the image denoising performance of the algorithm.

3.3 Noise classification model based on improved BPNN

In NLM, the filtering window is used to search the entire image area and identify noise pixels. Then the noise is separated to complete the image filtering. For the recognition of noise pixels, the main way is to estimate the weight of similar pixels. However, in the process of calculating the pixel weight, it is affected by image noise, which affects the recognition accuracy of noisy pixels. Therefore, BPNN model is proposed to fit and learn noise data and normal image data, so as to realize intelligent recognition and classification of noise points. The BPNN model is shown in Fig. 4.

Fig. 4

The structure of BPNN model.

For the image to be denoised, it is divided into overlapping neighborhood blocks, and each neighborhood block is denoised using the trained BP neural network model. In the NLM algorithm, weights are calculated based on the similarity of neighboring blocks, and the denoising results are weighted and averaged with adjacent pixels to obtain the final denoised image. The advantage of using BPNN to classify noise points is that it has higher efficiency and accuracy, and can realize large-scale image noise recognition, thus improving the image denoising efficiency and precision of NLM. The function of BPNN is largely related to the setting of its initial parameters. If the value of initial parameters is poor, the convergence, accuracy and learning efficiency of BPNN will be greatly affected. For this reason, the research proposes strategies to optimize and improve the BPNN to enhance the ability to noise. Particle swarm optimization (PSO) is a relatively common swarm intelligence optimization algorithm, and its application effect and scope have been verified. PSO is applied to find the best initial parameters of BPNN. However, PSO is prone to premature, resulting in poor global optimization performance. Therefore, the research proposes strategies to improve PSO. In the flight of PSO particles, a shrinkage factor is added as a constraint, so that particles in PSO are not affected by the flight speed limit. At this time, the particle velocity is calculated by Equation (13).

$\begin{matrix} V_{i}^{n + 1} = φ [V_{i}^{n} + C_{1} \times ran d_{1} () \times \\ (Pbes t_{i} - X_{i}^{n}) + C_{2} \times ran d_{2} () \times (Gbest - X_{i}^{n})] \end{matrix}$ (13)

In Equation (13), C₁, C₂ represents two learning factors, which have a direct impact on the flight speed of particles in the early and late stages respectively; rand₁ and rand₂ are random vectors; $V_{i}^{n}$ indicates the speed at which particles with the order of i in the population fly to particles with the order of n in the population; Pbest_i indicates the best position of the particles in the population with the order of i during the flight; Gbest indicates the best position where all particles in the population pass during flight; φ refers to the contraction factor proposed by the study, which can adjust the flight speed of all particles in the population in the way shown in Equation (14). $φ = \frac{2}{| 2 - C^{'} - \sqrt{C^{2} - 4 C} |}, C^{'} = C_{1} + C_{2}, C > 4$ (14)

In addition, in view of the defect that the global optimization performance of PSO is not ideal, Levy flight mechanism is used to optimize the relationship weight of PSO. Through this strategy, the search range and trajectory of the population are optimized, thus improving the global optimization ability of PSO. Based on the above contents, the optimized PSO (IPSO) is constructed experimentally. It is applied to the optimization of BPNN to build IPSO-BPNN model, as shown in Fig. 5.

Fig. 5

IPSO-BPNN model.

In Fig. 5, the network architecture of BPNN is first determined, and then the PSO algorithm is applied to optimize the parameters of BPNN. The optimized network randomly generates an initial particle swarm and calculates fitness values. If the conditions are not satisfied, the individual optimal solution and the group optimal solution are determined, and the shrinkage factor and Levy flight mechanism are introduced to adjust the particle’s flight speed and inertia weight. If the conditions are met, the error is calculated for each particle and the network parameters of the BPNN are updated. Until the final constraint conditions are met, the output is obtained and the optimal solution is found. Finally, combined with the above content, the IPSO-BPNN-IWOA-NLM (IBINLM) image denoising algorithm model is constructed experimentally. This model aims to achieve efficient and accurate image denoising, thus improving the quality of VCD.

4 Performance analysis of IBINLM image denoising model

A comparative experiment is designed to verify the performance of IBINLM image denoising model. The experimental environment is as follows. The system selects Windows 10 64bit and the memory is 8GB. Collect high-quality image data from the internet to build an experimental sample set. According to the ratio of 70% and 30%, the sample data set consists of training sample set and test sample set. Gaussian white noise of 10, 15, 20, 25 and 30 is added to the sample set respectively. The image after adding noise is the image that needs to be processed in the experiment. First, the recognition and classification ability of IPSO-BPNN to noise is tested.

In the training process, the error value and loss value changes of IPSO-BPNN model, PSO-BPNN model and BPNN model are compared to verify the optimization effect of IPSO on BPNN model. In Fig. 6, in the initial stage, as the number of iterations increases, the error and loss values of the model rapidly decrease. After a certain degree of decline, the error and loss values of the model gradually slow down and eventually no longer change. This indicates that after multiple iterations and updates, the performance of the model has reached its optimal state, and the error and loss values have been minimized. When the number of iterations is 10, 30, 50, and 100, the error values of the IPSO-BPNN model are 0.36, 0.12, 0.08, and 0.03, respectively. The error values of the PSO-BPNN model are 0.50, 0.28, 0.22, and 0.05, respectively. The error values of the BPNN model are 0.55, 0.35, 0.30, and 0.25, respectively. The loss values of the IPSO-BPNN model are 0.9, 0.19, 0.16, and 0.08, respectively. The loss values of the PSO-BPNN model are 1.11, 0.85, 0.79, and 0.21, respectively. The loss values of the BPNN model are 1.24, 1.00, 0.88, and 0.81, respectively. The above results can show that the efficiency and precision of IPSO-BPNN model have been significantly improved after optimization, indicating that the optimization effect is better.

Fig. 6

Optimization effect of IPSO on BPNN.

To test the denoising effect of the model, the deep convolutional neural network model (DCNN) and the support vector machine model combined with principal component analysis (KPCA-SVM) are compared with the IPSO-BPNN model, and the F1 value and accuracy are used as test indicators. Compare the effects of IPSO-BPNN model, KPCA-SVM model and DCNN model in image noise pixel recognition, as shown in Fig. 7. In Fig. 7, the recognition accuracy of noise pixels by IPSO-BPNN model (98.64%) is 5.33% and 6.58% higher than that by KPCA-SVM model and DCNN model, respectively. In Fig. 7(b), the F1 value of IPSO-BPNN model (96.32%) is 4.02% and 6.18% higher than that of KPCA-SVM model and DCNN model, respectively. Therefore, the IPSO-BPNN model has better detection performance.

Fig. 7

The effect of model in image noise pixel recognition.

To test the classification performance of the model, the true values are used as fitting lines, and the detection values of the IPSO-BPNN model, KPCA-SVM model, and DCNN model are used as data to analyze the fit of the three models. In Fig. 8, the fitting degree of IPSO-BPNN model, KPCA-SVM model and DCNN model to the input data is shown. The fit degree of IPSO-BPNN model (0.983) is 0.025 and 0.041 higher than that of KPCA-SVM model and DCNN model, respectively. The above results verify the application effect of IPSO-BPNN model in image noise pixel recognition and classification.

Fig. 8

Fit of model.

To verify the denoising effect of the IBINLM algorithm, two images are selected in the sample set. The additive Gaussian white noise with noise standard deviation of 20 is added to the image, and the image is denoised by IBINLM algorithm, NLM algorithm and wavelet transform algorithm (WT). Compare the image denoising effect in Fig. 9. IBINLM algorithm has the best function in image denoising, which can effectively remove image noise and retain the details of the original image.

Fig. 9

Image denoising effect of different algorithms.

Taking the Peak to Signal Ratio (PSNR) of the image as the detection index, PSNR of IBINLM algorithm, NLM algorithm and WT algorithm under different noise standard deviations is compared, as shown in Table 1. In different noise standard deviations and different images, the average denoising PSNR of IBINLM algorithm is 35.86dB, 2.09dB and 3.32dB higher than that of NLM algorithm and WT algorithm respectively. Therefore, the better the signal quality processes by the IBINLM algorithm, the smaller the distortion.

Table 1

PSNR under different noise standards

Image	Noise standard deviation	PSNR/dB
		IBINLM	NLM	WT
Woman	10	38.42	36.08	34.44
	15	36.85	34.42	33.46
	20	35.63	33.18	32.07
	25	33.58	32.02	30.17
	30	31.43	29.08	28.14
Man	10	40.33	38.42	37.85
	15	38.65	36.33	35.42
	20	36.70	34.58	32.98
	25	34.02	32.19	30.86
	30	32.95	31.44	30.02
Average	10	35.86	33.77	32.54
Ranking	–	1	2	3

The denoising MSE of IBINLM algorithm, NLM algorithm and WT algorithm under different noise standard deviations is shown in Table 2. Under different noise standard deviations and different images, the average denoising MSE of BINLM algorithm is 0.29, which is 0.23 and 0.38 lower than that of NLM algorithm and WT algorithm respectively. The BINLM algorithm has better denoising performance.

Table 2

MSE under different noise standards

Image	Noise standard deviation	MSE
		IBINLM	NLM	WT
Woman	10	0.07	0.14	0.23
	15	0.16	0.25	0.38
	20	0.27	0.48	0.55
	25	0.35	0.64	0.73
	30	0.56	0.85	0.96
Man	10	0.10	0.16	0.26
	15	0.17	0.36	0.53
	20	0.29	0.57	0.74
	25	0.38	0.72	1.06
	30	0.52	0.98	1.23
Average	10	0.29	0.52	0.67
Ranking	-	1	2	3

The ROC curves of IBINLM algorithm, NLM algorithm and WT algorithm are shown in Fig. 10. The AUC value of IBINLM algorithm reaches 0.903, which is 0.012 and 0.020 higher than that of NLM algorithm and WT algorithm respectively. The above results show that the image denoising ability of IBINLM algorithm is superior to the existing image denoising methods. This technique can successfully eliminate the image’s disturbance, and improve the quality and effect of VCD.

Fig. 10

ROC curve of algorithm.

The running time of IBINLM algorithm, NLM algorithm, and WT algorithm is solved using the large O representation method, and the results are shown in Fig. 11. The running time of IBINLM algorithm, NLM algorithm, and WT algorithm is solved using the large O representation method, and the results are shown in Fig. 11. In Fig. 11, the more samples, the faster the IBINLM algorithm runs. When the number of samples is 100, the running time of the IBINLM algorithm is only 12 seconds, which is significantly better than the other two algorithms.

Fig. 11

Time comparison of three algorithms.

In summary, the IPSO-BPNN model has significantly improved efficiency and accuracy compared to the PSO-BPNN model and BPNN model, indicating better optimization results. The IPSO-BPNN model has a recognition accuracy of 98.64% for noisy pixels, which is 5.33% and 6.58% higher than KPCA-SVM model and DCNN model, respectively. In Fig. 7(b), the F1 value of the IPSO-BPNN model reaches 96.32%, which is 4.02% and 6.18% higher than the KPCA-SVM model and DCNN model, respectively. The fitting degree of the IPSO-BPNN model reaches 0.983, which is 0.025 and 0.041 higher than the KPCA-SVM model and DCNN model, respectively, verifying the application effect of the IPSO-BPNN model in image noise pixel recognition and classification. In different noise standard deviations and images, the average denoising PSNR of the IBINLM algorithm is 35.86dB, which is 2.09dB and 3.32dB higher than the NLM algorithm and WT algorithm, respectively. Therefore, the better the signal quality processed by the IBINLM algorithm, the smaller the distortion. In different noise standard deviations and images, the average denoising MSE of the BINLM algorithm is 0.29, which is 0.23 and 0.38 lower than the NLM algorithm and WT algorithm, respectively. The BINLM algorithm has better denoising performance. The AUC value of IBINLM algorithm reaches 0.903, which is 0.012 and 0.020 higher than NLM algorithm and WT algorithm, respectively. The above results indicate that the IBINLM algorithm has better image denoising ability than existing image denoising methods.

5 Conclusion

In VCD, noise is easy to appear, which affects the quality and effect of VCD works. Therefore, based on depth learning and NLM, an IBINLM algorithm is proposed to achieve efficient image denoising. The experimental results show that when the number of iterations is 10, 30, 50, and 100, the error values of the IPSO-BPNN model are 0.36, 0.12, 0.08, and 0.03, respectively. The error values of the PSO-BPNN model are 0.50, 0.28, 0.22, and 0.05, respectively. The error values of the BPNN model are 0.55, 0.35, 0.30, and 0.25, respectively. The loss values of the IPSO-BPNN model are 0.9, 0.19, 0.16, and 0.08, respectively. The loss values of the PSO-BPNN model are 1.11, 0.85, 0.79, and 0.21, respectively. The loss values of the BPNN model are 1.24, 1.00, 0.88, and 0.81, respectively. IPSO-BPNN model has a recognition accuracy of 98.64% for noise pixels, which is 5.33% and 6.58% higher than KPCA-SVM model and DCNN model respectively. IPSO-BPNN’s F1 value reaches 96.32%, which is 4.02% and 6.18% higher than KPCA-SVM model and DCNN model respectively. The fitting degree of IPSO-BPNN model reaches 0.983, which is 0.025 and 0.041 higher than KPCA-SVM model and DCNN model respectively. It shows that the IPSO-BPNN model has good noise recognition effect. The average denoising PSNR of IBINLM algorithm is 35.86dB, which is 2.09dB and 3.32dB better than NLM algorithm and WT algorithm respectively. MSE is 0.29, which is 0.23 and 0.38 lower than NLM algorithm and WT algorithm respectively. AUC value reaches 0.903, which is 0.012 and 0.020 higher than NLM algorithm and WT algorithm respectively. The above results show that the image denoising ability of IBINLM algorithm is superior to the existing image denoising methods, which can successfully eliminate picture disturbance and enhance the quality and effect of VCD. Using the IBINLM algorithm to remove visual noise can make information clearer and help people better understand and interpret the design information and intention, improving the visual experience and comfort of the audience. Meanwhile, the IBINLM algorithm can be applied to multiple design fields to improve the clarity, attractiveness, and communication effect of the design. By removing visual noise, the design can better communicate its message and provide a better user experience. There are shortcomings in this study, as the data samples used are relatively small. In future work, it is required to increase the number of samples to improve the generalization ability of the algorithm.

Footnotes

Declarations

References

Shi

, Tang

and Yang

, Hyperspectral image denoising using a 3-D attention denoising network, IEEE Transactions on Geoscience and Remote Sensing 59(12) (2021), 10348–10363.

Parizi

M.K.

, Keynia

and Bardsiri

A.K.

, HSCWMA: A new hybrid SCA-WMA algorithm for solving optimization problems, International Journal of Information Technology & Decision Making 20(02) (2021), 775–808.

Zhang

, Xiao

and Tian

, A robust deformed convolutional neural network (CNN) for image denoising, CAAI Transactions on Intelligence Technology 8(2) (2023), 331–342.

Zhang

, Li

and Parizi

M.K.

, HWMWOA: A Hybrid WMA– WOA algorithm with adaptive cauchy mutation for global optimization and data classification, International Journal of Information Technology & Decision Making 22(04) (2023), 1195–1252.

Yan

, Li

and Zhang

, Depth image denoising using nuclear norm and learning graph model, ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM) 16(4) (2020), 1–17.

Liu

and Zhang

, A novel denoising algorithm based on wavelet and non-local moment mean filtering, Electronics 12(6) (2023), 1461–1461.

Diallo

, Hu

and Li

, Deep embedding clustering based on contractive autoencoder, Neurocomputing 433(14) (2021), 96–107.

Gurrola-Ramos

, Dalmau

and Alarcón

T.E.

, A residual dense u-net neural network for image denoising, IEEE Access 9 (2021), 31742–31754.

Shi

, Tang

, Yang

, Liu

and Zhang

L.P.

, Hyperspectral image denoising using a 3-D attention denoising network, IEEE Transactions on Geoscience and Remote Sensing 59(12) (2021), 10348–10363.

10.

Xie

, Wang

and Ji

, Noise2Same: Optimizing a self-supervised bound for image denoising, Advances in Neural Information Processing Systems 33 (2020), 20320–20330.

11.

Valsesia

, Fracastoro

and Magli

, Deep graph-convolutional image denoising, IEEE Transactions on Image Processing 29 (2020), 8226–8237.

12.

Zhang

, Li

, Sun

, Kong

and Fu

, Accurate and fast image denoising via attention guided scaling, IEEE Transactions on Image Processing 30 (2021), 6255–6265.

13.

Huang

J.J.

and Dragotti

P.L.

, WINNet: Wavelet-inspired invertible network for image denoising, IEEE Transactions on Image Processing 31 (2022), 4377–4392.

14.

Ezugwu

A.E.

, Agushaka

J.O.

and Abualigah

, Prairie dog optimization algorithm, Neural Computing and Applications 34(22) (2022), 20017–20065.

15.

Agushaka

J.O.

, Ezugwu

A.E.

and Olaide

O.N.

, Improved dwarf mongoose optimization for constrained engineering design problems, Journal of Bionic Engineering 20(3) (2023), 1263–1295.

16.

Agushaka

J.O.

, Ezugwu

A.E.

and Abualigah

, Gazelle optimization algorithm: a novel nature-inspired metaheuristic optimizer, Neural Computing and Applications 35(5) (2023), 4099–4131.

17.

Chan

K.Y.

, Kwong

C.K.

, Wongthongtham

, Jiang

H.M.

, Fung

C.K.Y.

, Abu-Salih

, Liu

Z.X.

, Wong

T.C.

and Jainc

, Affective design using machine learning: a survey and its prospect of conjoining big data, International Journal of Computer Integrated Manufacturing 33(7) (2020), 645–669.

18.

Aggarwal

, Mijwil

M.M.

, Al-Mistarehi

A.H.

, Alomari

, Gök

, Alaabdin

A.M. Zein

and Abdulrhman

S.H.

, Has the future started? The current growth of artificial intelligence, machine learning, and deep learning, Iraqi Journal for Computer Science and Mathematics 3(1) (2022), 115–123.

19.

Stojadinovi

, Kovaevi

, Marinkovi

and Stojadinović

, Rapid earthquake loss assessment based on machine learning and representative sampling, Earthquake Spectra 38(1) (2022), 152–177.

20.

and Sun

, Machine learning and AI in marketing-Connecting computing power to human insights, International Journal of Research in Marketing 37(3) (2020), 481–504.

21.

Meuwly

, Machine learning for chemical reactions, Chemical Reviews 121(16) (2021), 10218–10239.

22.

Parizi

M. Karimzadeh

, Keynia

and Bardsiri

A. Khatibi

, Woodpecker Mating Algorithm (WMA): a nature-inspired algorithm for solving optimization problems, International Journal of Nonlinear Analysis and Applications 11(1) (2020), 137–157.

23.

Parizi

M.K.

, Bardsiri

F. Keynia

and Khatibi

Amid

, OWMA: An improved self-regulatory woodpecker mating algorithm using opposition-based learning and allocation of local memory for solving optimization problems, Journal of Intelligent & Fuzzy Systems 40(1) (2021), 919–946.

24.

Parizi

M. Karimzadeh

, Keynia

and Bardsiri

A. Khatibi

, Woodpecker mating algorithm for optimal economic load dispatch in a power system with conventional generators, International Journal of Industrial Electronics Control and Optimization 4(2) (2021), 221–234.

25.

Gong

and Karimzadeh

, Parizi, GWMA: the parallel implementation of woodpecker mating algorithm on the GPU, Journal of the Chinese Institute of Engineers 45(6) (2022), 556–568.

26.

Zhong

, Wen

, Ma

et al. A hierarchical multi-leadership sine cosine algorithm to dissolving global optimization and data classification: The COVID-19 case study, Computers in Biology and Medicine 164(12) (2023), 107212–107212.