A Novel Medical Image Fusion Method Using Multi-Channel Pulse Coupled Neural Networks

A novel method for multi-channel pulse coupled neural networks (PCNN) is proposed in order to overcome the difficulty in lack of M-PCNN and question of many images fusion and instability of double-channel PCNN due to the different fusion sequence. In this study a novel multi-channel pulse coupled neural networks is analyzed in detail from odd numbers channel PCNN and even numbers channel PCNN. Through a combination of odd numbers channel PCNN and even numbers channel PCNN this study can realize image fusion by M-PCNN. A novel calculation method about linking-weight based on gray energy is proposed in order to overcome the lack of linking-weight in M-PCNN. To effectively avoid the Laplasse operator and calculation repeatedly, this study can realize not only many images fusion, but also better time efficiency and less information loss. Experimental results show that the presented method outperforms existing methods, in both fusion effect and different performance evaluation criteria.


I. INTRODUCTION
Pulse coupled neural network (PCNN) is the fourth generation of neural network model based on the pulse mechanism of the mammalian visual cortex. It is widely used in the field of multi-focus images, medical image fusion and image segmentation for its simple calculation, fewer parameters and good practicability. With the gradual deepening of research on single-and dual-channel PCNN, the method is becoming one of the most popular method in the image fusion. Compared with the wavelet transform [1] that has been applied to multi-focus image fusion, FSD [2] and Gradient pyramid [3], PCNN is still a research focus of multi-focus image fusion, medical image fusion, and the well-known works are [4]- [20].
Firstly, the relevant scholars put forward and summarized the background, principles, further development of the condition and application prospects of PCNN model [4]- [10], which laid the foundation for the further development of the model. Then, an adaptive dual-channel pulsecoupled neural network (PCNN) with triple-linking strength (ATD-PCNN) is proposed [11]. A novel fusion algorithm based on the adaptive dual-channel unit-linking pulse coupled neural network (PCNN) for infrared and visible images fusion The associate editor coordinating the review of this manuscript and approving it for publication was Yizhang Jiang . in nonsubsampled contourlet transform (NSCT) domain is proposed in [12]. The flexible multi-resolution and directional expansion for images of NSCT are associated, beacause of the global coupling and pulse synchronization characteristic of dual-PCNN. This is the biggest advantage of this method in medical image fusion research, a fusion rule [13] based on region spatial frequency is adopted in low frequency sub-band coefficient. Dual-channel PCNN has a simpler network architecture and better adaptability. But spectral distortion and excessive introduction of noise in fused images are often a common problem. To address this issue, a dual-channel SAR-optical images fusion algorithm based on non-subsampled shearlet transform (NSST) was adopted in Literatures [14], after NSST decomposition is performed. Then, a weighted average fusion rule using the coefficient of variation according to the different imaging characteristics of SAR images in the low-frequency sub-band of the decomposition image, the effect of SAR image noise on the fused image is removed by setting the coherence coefficient threshold in the high-frequency sub-band of the decomposition image. Because the calculations in prior research are complex, model in [15] has been further improved. Side-Scan Sonar Image Fusion Based on Sum-Modified Laplacian Energy Filtering and Improved Dual-Channel Impulse Neural Network is obtained, An improved PCNN model is proposed in this paper. This is the study's greatest contribution. Many scholars give their opinions on this question which greatly improved the computational complexity and efficiency. There are many scholars concerned about the combination of PCNN model and other models, which can foster strengths and circumvent weaknesses, and achieve the best fusion effect. Such as the combination of wavelet transform and other networks is proposed in the literature [16], [17]. We first decompose the infrared and visible image into multi-direction and multi-scale coefficient matrices by RLNSST decomposition. Then, Laplacian filter and Gaussian filter is utilized to generate the saliency map. Next, a novel weighted average based on guided filtering to retain edge detail information is used as the fusion rule. Finally, inverse RLNSST transform is used to get the fused image [16]. On this basis, a new fusion image fusion method is proposed in literature [17] based on multi-level Gaussian curvature filtering (MLGCF). We first decompose the input source images into three different layers:(small-scale, large-scale, and base layers). Then, three fusion strate giesare applied to combine the three types of layers. Finally, fusion image is reconstructed by summing the three types of fused layers. In literature [18], a novel fusion framework for infrared and visual images based on a full convolutional network (FCN) in the local non-subsampled shearlet transform (LNSST) domain is proposed. In other areas of research, such as literature [19], Visual Attention Technique (VAT) is mentioned. In PCNN models research, because both the two kind of image fusion methods above need to input the model for operations, which will bring the problem of low efficiency. To address this issue, a dual-channel PCNN model was adopted in Literatures [20], [21], using two channels to input information of two images at the same time to complete the fusion. Because the calculations in prior research are complex, model in [22] has been further improved. A new dual-channel PCNN model is obtained, with adding operation instead of multiplication, which greatly improved the computational complexity and efficiency. There are many scholars concerned about the combination of PCNN model and other models, which can foster strengths and circumvent weaknesses, and achieve the best fusion effect. Such as the combination of wavelet transform and PCNN is proposed in the literature [23]- [25]. LSWT image (Lifting Stationary Wavelet Transform) is decomposed into high and low frequency parts, using PCNN model for fusion in the high-frequency part [23]. On this basis, a new fusion image fusion method is proposed in literature [24], [25] based on non-subsampled contourlet transform (NSCT) and adaptive unit-fast-linking PCNN. This method not only improved the PCNN model, synthesizing the advantages of both unit-linking PCNN and fast-linking PCNN, but also took advantage of the clear pixels in the image as a linking weight. A image fusion method is presented in [26] by combining contourlet hidden markov tree (CHMT) and PCNN model. Firstly image is decomposed by CHWT, the rules that got the largest fusion was adopted in the low-frequency part, PCNN model was used for integration in the high frequency part, links coefficient in model was determined by the clarity of image. This fusion method has been successfully used in remote sensing image fusion.
Image fusion method based on PCNN model has been successfully used in many areas. However, these PCNN-based algorithms are limited to dual-channel PCNN model. With the development of the application, the methods above are facing some common challenges: one of the key challenges is the methods above are only designed for two images fusion, and some problems will appear when facing the synchronous fusion of multiple images.
With the continuous advancement of technology and development of society, the problem that we need to fuse multiple images into a more informational and clear image will appear in many fields. For example, monitoring probes will record a large number of images per day, how to integrate the distributed information on multiple images into a more valuable image, is gradually becoming a new problem be solved. At present the existing PCNN models' detailed practice to solve this problem is: two images are fused firstly, and then the third image is fused with the fusion image, in turn until the final fusion image is obtained. The problems according to this method to achieve the fusion of multiple images are as follows: (1) It's difficult to determine the order of multiple images to be fused, and different order lead to different results.
(2) A high time cost is paid.
(3) The linking weights and other information are determined by the condition of two images when they are fused. Due to lack of global considerations, without thinking about the situation of all images which will cause the loss of some valuable information.
In order to solve these problems more effectively, and for the research on multi-channel PCNN model is rarely involved recently, in this paper, a detailed study of the multi-channel image fusion model has been carried out. The major contributions of this paper are as follows: (1) We proposed a multi-channel PCNN image fusion model to achieve the synchronous fusion of multiple images.
(2) In order to achieve a better process of the multi-channel PCNN image fusion, a suitable method for calculation of the linking weight of multi-channel image fusion model is proposed in this paper. It is not only solves the key challenge in multi-channel PCNN image fusion method, but also contacts with the gray energy of image closely, which reduces the loss of information, enhances the fusion efficiency and reduces the fusion time.
(3) The multi-channel PCNN image fusion algorithm is proposed on the basis of previous conclusions and analysis.
The rest of this paper is organized as follows. In Section 2, the multi-channel PCNN image fusion model is briefly reviewed, and then the improved method of weight calculation is introduced in Section 3. The multi-channel PCNN image fusion model algorithm from two angles of the odd channels and even channels is proposed in Section 4. Experimental results and performance evaluation are given in Section 5. Conclusions are drawn in Section 6.

II. MULTI-CHANNEL PCNN MODEL A. A SINGLE-AND DOUBLE-CHANNEL PCNN MODELS
A brief review for the classic single-and double-channel PCNN models are given in this section. As the model shown in Fig.1, feeding input and linking input as the input components are used for input, stimulation and feedback, resulting in a series of ignition iterative process. The model is first applied to image fusion, which was proposed in literature [11]. Two images input into model respectively for operations, then take the larger ignition times on the corresponding pixel of the two respective ignition matrix as a fusion result, which will inevitably bring the problem of computational complexity, time efficiency is low, the fusion accuracy is not high. For such problems, a improved double-channel PCNN model was proposed in literature [12], [13]. In the improved model shown in Fig.2, the feeding input and linking input are as the two equal branches to input the information of two images synchronously, and the fusion process is obtained in fusion pool after a series of operations.

B. MULTI-CHANNEL PCNN MODEL
Although the dual-channel PCNN model greatly improved the defects of the complex calculation and low accuracy in single-channel PCNN model, two input channels are not enough. At present, few researches of multiple images synchronous fusion are presented. The main idea of existing method is as follows: two source images input dual-channel PCNN model to complete the fusion firstly, and then the third image is fused with the fusion image, in turn until the final fusion image is obtained. However, according to the results of our study we found that different fusion order of multiple images would have a great impact on the fusion result. Especially the choose of first a few images is very important, if the fusion result is not ideal in the beginning, it is bound to have a great adverse effect on the final fusion result. Furthermore, it is a waste of time to achieve multiple images fusion by using dual-channel PCNN model. From the above consideration, a multi-channel PCNN model is proposed in this paper, which is better suited to the  Fig.3. Fig.3 shows a multi-channel PCNN model on the basis of double-channel PCNN model by adding to N input branches, and the number of corresponding linking weighted branch is N, but the mechanism of pulse generation and ignition occurred is unchanged. In order to further improve the computational efficiency of the multi-channel PCNN model, the proposed algorithm has simplified the multiplication in Fig.2, by combining with other multiplication in fusion pool, and only the addition operator is obtained in internal output. The following expressions describe its mathematical model: where H A , . . . . . . , H N refer to N images are input, refers to the output of model, β A , . . . . . . , β N refer to the weighting coefficients of N channels, σ is a adjustable parameter to adjust the average level of the internal output, refers to the number of ignition, α T is the decay time.

III. THE CALCULATION OF LINK-WEIGHT A. THE CALCULATION AND PROBLEM OF LINK-WEIGHT IN DUAL-CHANNEL PCNN MODEL
Linking weight β presents the strength of the neurons' linking actions in the PCNN model. β = 0 presents all neurons are independent and non-interfering. β = 0 refers to the existence of coupling between neurons, which means a neuron has fired, adjacent neurons may lead to be fired, and β determines the scope of the adjacent neurons. The larger β is, the greater the area of synchronous pulse method and captured neurons brightness. On the contrary, the smaller β is, the fewer images details are obtained, which means we can only achieve some coarse information of framework. In this paper, each pixel in the image gray matrix is as a neuron in the double-channel PCNN model. Let β be a large value, so that more images' information in this channel can be obtained. Therefore, β is regarded as the weight parameter between all channels in the double-channel PCNN model, and the parameter has an important impact in fusion effects.
At present, the methods to calculate the link-weight are average gradient method, Energy of Laplacian of Image [12] (EOL) and Sum Modified Laplacian [13] (SML). But all these methods have their shortcomings. For example, average gradient method gets related calculation only from the clarity of images, which is easy to overlook such as the brightness, the amount of information contained and so on. EOL method exists the problem of duplicate calculations. SML method improves Laplacian operator template on the basis of EOL method, and uses the sliding window technique to achieve the summation within local scale. But SML method still has the problem of computational complexity and duplicate calculations, in the meanwhile, the result of this method has a closely relation with the selection of Laplacian operator template and weight matrix. In addition, image energy and Laplacian operator template need to be determined firstly when computing link-weight, which introduced a lot of time for the fusion process. The methods above have demonstrated some validity in double-channel PCNN model to some extent, but they do not apply to multi-channel image fusion model.
To solve the problem above, we recalculate the image energy based on image gray matrix, and then determine the link-weight. The proposed algorithm has solved the determination of the link-weight in the multi-channel PCNN model, and has also improved the computing speed effectively with guaranteeing the accuracy of the premise.

B. IMPROVED METHOD OF THE LINKING CALCULATION
Because the gray level matrix of image can reflect the characteristics of the image in a way, such as the value of each VOLUME 8, 2020 pixel in the image matrix can reflect the brightness of the image, usually the greater the gray value, the greater the brightness, and the brightness is the energy of an image. Replacing EOL by the gray matrix can not only obtain the energy information, but improve the computation efficiency. In this way, the information on the focus energy of the image can be computed more precisely with low computation cost.
As shown in Eq.(6)- (7), the improved calculation method of link-weight calculated for the gray matrix of image, and the weight index of two branch links set as positive and negative respectively, which will achieve better overall coupling.

C. THE CALCULATION OF LINK-WEIGHT IN MULTI-CHANNEL PCNN IMAGE FUSION MODEL 1) THE CALCULATION OF MULTI-CHANNEL LINK-WEIGHT
Based on the above analysis, we proposed a new calculation method of link-weight in multi-channel PCNN model which is more accurate and simple, see Eq. (8)- (9).
where N presents the number of channels, S A ij . . . . . . S N ij is the Nth-channel input at pixel(i, j), η is adjustable parameter. In order to couple the information in each channel effectively, factor (−1) θα is introduced to achieve positive and negative coupling. In the meanwhile, only when the sum weight of each channel is close to 1, a better overall coupling is obtained. Then we set the denominator parts in Eq. (8) and (9) be multiplied by the coefficient γ , which is used to adjust the order of magnitude of the weight in each channel to meet the overall coupling much better.
The advantages of proposed method to calculate the link-weight are as follows: (1) By the adjustment of the denominator, the weights of all the channels are coupled and the main branch is highlighted. In this way, the effect of multiple branches can be guaranteed.
(2) Due to the direct use of the image gray level matrix instead of the results of calculations using a sliding window. The proposed method avoids the Laplacian operator template and sliding window calculation, it can greatly simplify the calculation process in ensuring the accurate calculation of the premise.
(3) A new calculation method for link-weight which applies to multi-channel image fusion model is proposed in this paper. The method lays the foundation for the further study of multi-channel image fusion. Fig.3 shows the structure of model and Eq. (6)-(7) present the link-weight calculation method of multi-channel PCNN model. From which we find the weight of each channel in even-number channels model just shows the weight of an even number of positive and negative coupling, and a channel will can't couple if it encounters the odd-channel model. Therefore, model in both cases of odd and even channels will show the different fusion result. In this paper, we have studied the effectiveness, time efficiency and accuracy of the proposed model for image fusion from two angles of odd and even channels PCNN model.

2) ANALYSIS OF LINK-WEIGHT CALCULATION METHOD
In addition, we found under realistic conditions, whether N is odd or even, the current methods can't solve when N → ∞. However, we can refer to the method of solving prime numbers in mathematics. In order to tackle multiple channel problems, a relatively simple model with fewer channel numbers are firstly investigated. The number of channels are often taken as a prime number. After that, the results are combined into the more complicated model. Because prime numbers less than 10 are 2, 3, 5 and 7, 5 is represented by 2 + 3 in this paper. Although 7 can be represented by 2 + 5, 5 can only be represented by a multiple of 5, and which covers a smaller range. In summary, 5-channel model will not be emphasis in the proposed method. If 7 is represented by 2 + 3 + 2, which is also not the emphasis in our study. Firstly, the study of double-channel PCNN model is already very mature. What's more, the defects of double-channel PCNN model for multi-channel fusion such as fusion results are influenced by the fusion order vulnerably, fusion accuracy is low, time-consuming and so on are still problems need to be solved. In this paper, we let 7 be represented by 3 + 4.

B. M-PCNN
The process of M-PCNN is: firstly, the gay matrix of image is as model input, and which should be normalized. Then, the linking weight of each channel is calculated by Eq. (8)-Eq. (9). Finally, the fused image is obtained by the VOLUME 8, 2020

M-PCNN
Step1: Initialization all zero for initial state model does not produce pulses.
Step4 (1) Synchronous fusion of multiple images can be more effective.
(2) Better to save fusion time.
(3) Maximize the good collaboration and lead to ignition from a global point of view for each channel uses reasonable and suitable method for multi-channel model to calculate weight.

V. EXPERIMENT AND ANALYSIS
To evaluate the performance of our proposed fusion method, extensive experiments with multi-focus image fusion have been performed. Note that all test images used in the experiment are downloaded on the website (http://www.imagefusion.org [20]) and literature [13].

A. EXPERIMENT OF ODD NUMBER CHANNEL PCNN IMAGE FUSION
As described in Section III, 5-channel PCNN model were studied in detail as the basis of the odd-channel model in this paper. Several experiments with multi-focus image fusion (one set of size 256 × 256, another set of 480 × 360) have been performed between dual channel PCNN model and the proposed model. Fusion process of the experiment in dual-channel model is as follows: Firstly, five images were arranged in accordance with the different order. Then, dual-channel PCNN image fusion method was performed on the top two images. Finally, the final fusion result was obtained using the same fusion method performed on the fusion result and the other source images. Fig.5-Fig.6 and Table 2-3 show the experiment result and Table 1 shows the parameter settings. Experimental datum were the average of the results obtained by running operation algorithms 100 times. The evaluation index used in the experiment: the mean (MEAN), variance (STD), entropy (H), average gradient (GRAD), time (TIME).

1) EXPERIMENT OF CT1 IMAGE
Experiment images are multi-focus images with the size of 256 × 256 pixel, Fig.5 and Table 2 respectively show the fusion results and comparison of results in different fusion order.

2) EXPERIMENT OF CT2 IMAGE
Experiment images are multi-focus images with the size of 480 × 360 pixel, Fig.6 and Table 3 respectively show the fusion results and comparison of results in different fusion order.

3) EXPERIMENTAL ANALYSIS a: TIME PERFORMANCE ANALYSIS
Time performance analysis is very important to evaluate the performance of a fusion method. In this section, we carried out a detailed comparative analysis for the above two sets of experiments from the time performance.
As shown in Fig.7, the proposed method has a better time performance than other methods. It spends only about VOLUME 8, 2020   two-thirds time than other methods, and a fusion results with greater clarity and resolution can be obtained.

b: VISUAL EFFECT AND OBJECTIVE EVALUATION INDEX ANALYSIS
In this subsection, we had a detailed analysis of above two groups of experimental results from fusion effect and performance evaluation. From the above two sets of experimental results, we found some differences between six fusion results of different fusion order. Because different fusion order result some differences in gray-scale image, clarity, resolution and time. For example, the fusion result in the first set of experiments which is obtained by dual channel model in the 1-2-3-4-5 fusion order has a maximum average gradient, but also more time-consuming. The fusion result in the second set of experiments which is obtained by dual channel model in the 1-2-3-4-5 fusion order has a maximum information entropy, but also more time-consuming. Fusion result according to the 4-1-2-3-5 fusion order has a minimum mean. However, the image obtained by the proposed method is significantly better than the image obtained by dual channel PCNN image fusion method in the three aspects of image mean (MEAN), variance (STD), the average gradient (GRAD). It just shows the image obtained by the proposed method has larger image contrast and higher resolution, image gray tends to disperse, and obviously reacts the tiny details contrast in the image. In this paper, a reasonable weight calculation method is adopted, each channel can   maximize the good collaboration, and information loss is less than the dual-channel pulse-coupled model. Compared to other methods, the proposed method can effectively solve the problem that different fusion orders result in differences of the fusion results. However, the proposed method is slightly worse than the dual-channel PCNN image fusion method in the aspect of information entropy (ENT).

c: IGNITION ANALYSIS
In order to further analyze the reason why 5-channel PCNN image fusion model can obtain a better result, we conducted a detailed analysis to the ignition situation caused by the 5-channel PCNN image fusion model.
The blue area in Fig.8 represents ignition area. It can be seen from two figures that the blue area in the figure is much more, which illustrate the most pixels can cause ignition phenomenon by the proposed method. Most importantly, it reflects the method to calculate the link-weight for each channel by gray-scale energy of image grayscale matrix is much better, the energy in each channel did not loss for mutual interference, and mutual fusion can maximize the ignition effect. On the other hand, good ignition situation illustrates that the ignition matrix can completely reflect grayscale energy information of the multiple images, which is bound to bring a better fusion effect.

B. EXPERIMENT OF EVEN NUMBER CHANNEL PCNN IMAGE FUSION
As described in Section III, 8-channel PCNN model were studied in detail as the basis of the even-channel model in this paper. Several experiments with multi-focus image fusion have been performed between dual channel PCNN model and the proposed model, and the size of multi-focus VOLUME 8, 2020 images are 256 × 256, 480 × 360.196 × 146. Fusion process of the experiment in proposed model is as follows: Firstly, arrange the four source images in a different order. Then, dual-channel PCNN image fusion method was performed on the top two images. Fig.9-Fig.11 and Table 5-Table 7 show  the experiment result and Table 4 shows the parameter settings. Experimental datum are the average of the results obtained by running algorithms 100 times. The evaluation index used in the experiment: the mean (MEAN), variance (STD), entropy (H), average gradient (GRAD), time (TIME).

1) EXPERIMENT OF CT1 IMAGE
Experiment images are multi-focus images with the size of 256 × 256 pixel, Fig.9 and Table 5 respectively show the fusion results and comparison of results in different fusion order.

2) EXPERIMENT OF CT2 IMAGE
Experiment images are multi-focus images with the size of 480 × 360 pixel, Fig.10 and Table 6 respectively show the fusion results and comparison of results in different fusion order.

3) EXPERIMENT OF CT2 IMAGE
Experiment images are multi-focus images with the size of 196 × 146 pixel, Fig.11 and Table 7 respectively show the fusion results and comparison of results in different fusion order.

4) EXPERIMENTAL ANALYSIS a: TIME PERFORMANCE ANALYSIS
In this section, we carried out a detailed comparative analysis for the above three sets of experiments from the time performance.
As shown in Fig.12, the proposed method has a better time performance than other methods.

b: VISUAL EFFECT AND OBJECTIVE EVALUATION INDEX ANALYSIS
In this subsection, we had a detailed analysis of above three groups of experimental results from fusion effect and performance evaluation. From the above three sets of experimental results, we found some differences between 12 fusion results of different fusion order. Because different fusion order of 8 source images result some differences in gray-scale image, clarity, resolution and time.   is time-consuming 4.222765 seconds, significantly higher than the other dual-channel fusion results 0.3-0.8 seconds. However, the fusion image obtained by the proposed method is significantly better than the fusion result obtained by dual channel PCNN image fusion method in the four aspects of image mean (MEAN), variance (STD), information entropy (ENT), the average gradient (GRAD). It just shows the image obtained by the proposed method has larger image contrast and higher resolution, image gray tends to disperse, and tiny details of the image contrast are obvious. In this paper, a reasonable weight calculation method is adopted, each channel can maximize the good collaboration, and information loss is less than the dual-channel pulse-coupled model. Compared to other methods, the proposed method can effectively solve the problem that different fusion orders result in differences of the fusion results. However, the proposed method is slightly worse than the dual-channel PCNN image fusion method in the aspect of information entropy (ENT).

c: IGNITION ANALYSIS
In order to further analyze the reason why 8-channel PCNN image fusion model can obtain a better result, we conducted a detailed analysis to the ignition situation caused by the 8-channel PCNN image fusion model.
The blue area in Fig.13 represents ignition area caused by 4-channel PCNN image fusion model. Figure 13 (a), (b), (c) are virtually all of the blue region, which illustrate the most pixels can cause ignition phenomenon by the proposed method. Most importantly, it reflects the method to calculate the link-weight for each channel by gray-scale energy of image grayscale matrix is much better, the energy in each channel did not loss for mutual interference, and mutual fusion can maximize the ignition effect. On the other hand, good ignition situation illustrates that the ignition matrix can completely reflect gray-scale energy information of the multiple images, which is bound to bring a better fusion effect. VOLUME 8, 2020    In addition to this, by comparing Figure 8 (a) and Figure 13 (a), Figure 8 (b) and Figure 13(b), it is found that the pixel points could not cause ignition in Fig.8 (a) and (b) had caused ignition in Fig.13 (a) and (b), which illustrate that ignition condition caused by 8-channel PCNN model is better than 5-channel PCNN model for the same input image. This further illustrate the proposed fusion method not only does not cause loss of information, but also can maximize the fusion capabilities through collaboration between various channels when the number of channels increases. From the ignition situation analysis above, we found the ignition situation is much better when the number of channels is less than or equal to 8. And which is consistent with experience in our daily lives that more raw data collected from more valuable information.
In short, the above analysis shows that the fusion images obtained by both the odd channel model and even channel model have larger image contrast and higher resolution, image gray tends to disperse, and tiny details of the image contrast are obvious. Which verifies the proposed method can achieve multiple images synchronous fusion with less time cost, and effectively solves the problem that different fusion order results in differences fusion results.
The advantages of this paper are as follows.
(1) As a result of a reasonable weight calculation method, each channel can maximize the good collaboration from a global perspective, the loss of valuable information is significantly less than the dual-channel PCNN model.
(3) The fusion image obtained by the proposed method has gray dispersion, large image contrast, high definition, high resolution in a certain extent.
(4) Because the proposed method is better linked with the image gray energy, fusion effect is better for images which are able to highlight the image gray information.
(5) Rarely involved in the plight of the multi-channel model has been effectively solved.

VI. CONCLUSION
In order to better solve the problems of synchronous fusion of multiple images and rarely involved in the plight of the multi-channel model, effectively overcome the problem that a great differences of result brought by dual channel PCNN model in different fusion order, a multi-channel PCNN image fusion method is proposed in this paper. Firstly, the multichannel PCNN image fusion model is proposed. Then, calculation method of link weight has been improved. A detailed analysis of the multi-channel model from two angles of the odd channel PCNN model and even channel PCNN model is carried out, it is found that N-channel PCNN image fusion can be achieved through a combination of odd channel and dual channel. Finally, a detailed comparison with dual-channel PCNN image fusion method and non-PCNN image fusion method is carried out, which further proves the proposed method can not only solve the simultaneous integration of multiple images, but also shows certain advantages in fusion effect and time performance.
YI LI (Member, IEEE) is currently a Researcher in image processing and pattern recognition, with a major in mathematics and computer science, having plenty of work experience in virtual reality and augmented reality projects, and is engaged in the application of computer medical image diagnosis. Her research application fields range widely from deep research fields to everyday lives.
JUNLI ZHAO was born in 1977. She received the Ph.D. degree in computer applied technology from Beijing Normal University, Beijing, in 2015. She was a Visiting Scholar with the State University of New York at Stony Brook. She is currently an Associate Professor and a Master Supervisor with the School of Data Science and Software Engineering, Qingdao University, Qingdao. She is currently engaged in computer graphics, computer vision, and virtual reality.