A Novel Image Denoising and Segmentation Using Machine Learning with SRM Strategy

Objectives: In the last decade, artiﬁcial intelligence (AI) and machine learning (ML) have a signiﬁcant impact on image analysis and segmentation. The most demanding aspect in digital image processing is eﬃcient segmentation to extract the desired features or recognize the hidden patterns of the digital images, which are one of the major challenges. Even in existing segmentation techniques, due to under segmentation issues, particularly in remote or arial image analysis, necessary object features and residual details are not properly segmented. Therefore, the study presents a new method for image segmentation using machine learning, which addresses these limitations. Methods: This article presents a novel approach for arial image denoising and segmentation using unsupervised learning. The Discrete Wavelet Transform (DWT) is primarily used as a

simulation results have proven to produce better segmentation and denoising with less iteration time. A qualitative evaluation through visual examination justifies the proposed study and is superior to some of the popular methods. Novelty/Applications: The utilization of wavelet transform to preserve edge features during segmentation is one of the key features of this method, and it significantly overcomes the under segmentation patterns. The combination of minimum-spanning tree clustering (MST) and statistical region merging (SRM) is the major strategic step in segmenting and detecting the essential object features of arial images.

Introduction
Remote or arial image analysis is one of the vital research area and plays a critical part in object detection and classification. Various image segmentation algorithms have been developed over the years, however, clustering algorithms have an essential role in the segmentation of digital images. Since unsupervised segmentation of images doesn't require prior knowledge, most of the clustering algorithms employ unsupervised segmentation. It recognizes the hidden patterns of the image to bring feasible outcomes (1)(2)(3) . Some clustering techniques try to separate regions based on similarity criteria, which look like the decomposition of the image into independent areas (4)(5)(6)(7)(8)(9) . However, few machine learning algorithms don't provide a specific number of clusters during the partitioning process. Hence they generate as many clusters depending on the scaling factor of the image. Due to the extreme level of the clustering process, sometimes it is necessary to merge the unwanted or meaningless regions in an image, which helps to reduce the over-segmentation of the image (10)(11)(12)(13)(14) . Various region merging techniques were employed in the last few decades, to overcome the over/under segmentation patterns (15)(16)(17) . Nevertheless, based on the literature review, the combination of wavelet and watershed-based techniques in pre-processing phase is a superior technique in existing methods, the post-processing phase includes similarity based region merging, boundary-based merging, hierarchical merging, dynamic region merging and statistical region merging, etc, all these techniques arises from different predefined threshold criteria to merge required part of the image and the method provides better-segmented results only for standard images (18)(19)(20)(21) . Image analysis and segmentation are essential needs in military and defense applications, which helps in recognizing the particular objects based on the requirements. The study mainly focused on Arial image denoising and segmentation. The existing methods provide better outcomes for standard image segmentation with the usage of discrete wavelet transform (DWT) at pre-processing level, and the region merging process is utilized at the post-processing level. But in the case of arial image, they lack consistency (22,23) . Therefore, the study suggests a new method, which is applicable for both normal as well as arial images. The simulation result of the conducted study is compared with some of the existing methods and is found to be significantly better in terms of denoising and segmentation. Visual inspection justifies the segmentation quality of the proposed method has effective outcome than existing methods, and it recognizes the object portion and provides meaningful image details at the wavelet restoration phase (24,25) .

Discrete wavelet transforms (DWT
In DWT, the digital image is divided into four sub-images as, A (approximation coefficient or A1 band) contains most of the image information and D https://www.indjst.org/ (detailed coefficients or H1, V1, D1 bands). However, detailed coefficients contain edge/texture details of the image along with residual noise. Here machine learning-based image denoising is employed, which selects the appropriate shrink filter to detailed coefficients based on some prior knowledge (2)(3)(4)(5)(6) . A1 band represents most of the low-frequency image features. Which is subjected to the pre-processing step and is further decomposed to level-2 based on requirement (13)(14)(15)(16) , as shown in Figure 1.

Fast NL-means filter
During the period of hypothetical design of the filter, the value of pixel 'i' is repaired, where NL (v) (i) is given by Where, v is the intensity function, and thus v(j) is the intensity at pixel j and w(i,j) the weight associated with v(j) in the reconstruction of pixel i. The unique description of the fast NL-means procedure reflects that every pixel can be connected. But virtually, the number of pixels taken into relation in the weighted mean is limited in a neighborhood. The likeness among i and j based on the correspondence of their native regions with mean Euclidean distance (6)(7)(8) . The general difficulty that arises in this process is the computational load with a neighborhood pixel grouping. Therefore, the purpose is to diminish the number of pixels taken into relation in the weighted mean and to evade recurring designs of the similar neighborhood (9)(10)(11)(12) .
where, z(i) is the normalization constant with z(i) = ∑ j w (i, j) and h acts as a filtering parameter. v (N i ) and v (N j )v (N i ) and v (N j ) is the similarity between the neighborhood pixel intensities. "Distance between v (N i ) and v (N j ) is a standard norm ||1 − ∥ a , convolved by Gaussian-kernel of standard-deviation 'a' , although measure of the distortion between pixel neighborhood intensities".

Minimum-spanning tree clustering
Clustering is the most popular and dynamic research area in statistics, pattern analysis, computer vision, and it aims to cluster similar objects into the same group, and the objects fitting to different clusters are dissimilar. The minimum spanning tree https://www.indjst.org/ (MST) clustering process can identify the segments of arbitrary shapes by separating the consistent edges in the image. The definition of inconsistent edges is a major issue that has to be addressed in all MST-based clustering algorithms. Sometimes, the inappropriate number of image clusters, like false edges and excessive boundaries, may lead to over-segmentation (13)(14)(15)(16) . A spanning tree is an acyclic sub-graph of a graph G, which contains all the vertices from G. The MST of a weighted graph is the minimum weight spanning tree of that graph. The cost of constructing an MST is given by, where m is the number of edges in the graph and n is the number of vertices. However, the Euclidean distance between the vertices and global as well as local densities of the vertices is defined by adjusting the variance in the Gaussian function. Correspondingly, two groups of K-cluster centers under different densities are achieved. Finally, partitioning the image clusters leads to many clusters, and the process is repeated until the optimized clusters are formed (17)(18)(19)(20) .
where, G = (V, E) be connected, nodes and edges associated with weighted graph are given by Such that e i is weighted connection among two nearer nodes (V | = n, and (E| = m The intensity of the pixel at position (x, y) with associated node value is v i = I x,y . The similarity among the two nearer pixels is calculated by adding weight w j to the edge e j , similarly, intensity difference among the nearer pixels forms the edge e j . The path P is a series of edges linking the two nodes in a graph G is given by Therefore, the minimum spanning tree of a graph G, and the weight w i connected with edge e i of E' is,

Soft threshold
A soft threshold is a pre-processing tool that estimates the optimal thresholding for sub-bands in the image. Here, imagedenoising is done by either of three methods, namely Sure shrink or Bayesian shrink, or Visual shrink, based on the type of image used in the process. Appropriate filtering is employed by the machine, based on some prior knowledge and the type of the image. It provides image smoothing and better edge preservation. Since filtering is done in the wavelet domain, the detailed coefficients of the decomposed image features are processed, and pixels with intensity values above the threshold values are modified. Subsequently, the shrinkage reduces the noise without distorting the required image features (2)(3)(4)(5)(6)(7) , Due to the dual image filtering process, the contamination of noise details is significantly suppressed, during the image reconstruction, and at the same time, sharp image features are still retained in the wavelet restoration phase (8)(9)(10) . The sub-band dependent threshold estimation is given by in every cycle,β value is calculated at k th cycle, L k is the extent of sub-band, and the noise difference σ 2 is valued by HH1 sub-band ) σ 2 is the sub-band standard-deviation.

Statistical region merging (SRM)
The SRM algorithm belongs to the group of region merging and growing techniques combined with geometric tests to select the merging of a particular region depending on statistical criteria. In the projected study, SRM is employed by considering the segmentation errors while dealing with the clustering process. It would not be possible to separate a limited set of regions during the clustering process. Therefore, these image regions are merged by choosing the suitable decisions in computation of SRM (22)(23)(24)(25) . The region of interest for merging image labels is based on homogeneity and similarity between the regions. https://www.indjst.org/ Correspondingly, merging criteria are assigned for creating native decisions, and dissimilarity measures between the pixels with the boundary of the two regions are considered. Both consistency and divergence procedures are utilized to select the merging criteria and are called statistical similarity tests (26,27) .
By considering the regions R ′ , R of the digital image IM, if it shares the similar consistency measure, merging criteria is given by where, |.| denotes the cardinality and b(R) is given by Based on the ascending order of the pixel and the difference in pixel intensities between the candidate regions considered for the merging process. The predicate is satisfied with good probability p ≥ 1 − N_for N merge trials, where, (N) is the number of merging tests and maximum value of N<2 for an image IM The numbers of regions are varied, based on image and its scaling factor (28) . This paper is organized as follows: The proposed method is discussed in section 2, and section 3 contains the flowchart of the suggested algorithm. The experimental results are given in section 4, and the conclusions are in section 5.

Proposed method
This paper develops a new image segmentation technique based on a machine learning platform. Here random noise is added to the standard arial image. The decomposed image is obtained from Discrete Wavelet Transform (DWT). During pre-processing stage level-2 decomposition is carried out, and approximation coefficients are subjected to a fast NL-means filter, which gives a smoothing effect while preserving edge/texture details. The usage of unsupervised learning for denoised images provides a better-segmented image. Here Minimum-Spanning Tree (MST) clustering technique is mainly used to track the complete object features, which gives more significance to full segmentation as compared to some of the popular segmentation techniques.

Algorithm steps
Input: standard Arial images are extracted from the SSIP database for analysis and segmentation.
Output: Segmented image. · A random noise is added to the test image, and the usage of Discrete Wavelet Transform (DWT) provides the decomposed image.
· To reduce the noise, the approximation sub-image is subjected to a fast NL-means filter. · A Minimum Spanning Tree (MST) clustering approach is assigned to partition the image details, which extract the hidden features of the image.
· The usage of Statistical Region Merging (SRM) helps to reduce the over-segmentation patterns of clustered images. · Compatible soft threshold filter is assigned based on the structural image criteria is applied by the machine itself with its prior learning · A combination of wavelet coefficients is displayed into full resolution at the wavelet restoration phase.

Results and discussions
In the proposed study, arial images are used for analysis. Here, noisy image is subjected to wavelet domain, a LL sub-image is treated with fast N-L means filter to achieve better denoising effect, and residual details are filtered automatically by considering https://www.indjst.org/    Table 1. However, 1014 segmented regions of the clustered image (7.1.06) display the robustness of segmentation, and the algorithm is efficient to track the remote objects more meaningful than other leading methods. In the merging procedure, a statistical method is used to eliminate the over-segmentation patterns to bring out 29 essential object features with residual details at level-2 decomposition is represented in Figure 3 and is one of the major novelty adopted in this work, which indicates that the system is effective while tracking the required object features in a meaningful way. Nevertheless, the usage of Discrete Curvelet Transform (DCT) preserves finer details of the image. However, the method is limited to curve and needle-shaped images, and Stationary Wavelet Transform (SWT) provides a better denoising effect due to the loss of spatial resolution and texture details. These techniques are only suitable for specific applications at the pre-processing stage. Thus DWT is utilized in the suggested algorithm, which preserves texture details and provides better spatial resolution, especially for Arial images. The PSNR values along with the time are the main criteria, in which the method excels as is evident from the results, are illustrated in Table 2, and Figure 10. The visual inspection also justifies the proposed study and provides better segmentation, especially for Arial images. Finally, the experimental results are compared with other leading methods https://www.indjst.org/ with the combination of Hierarchical Region Merging (HRM) as well as Statistical Region Merging (SRM) techniques (Table 2, Figure 11).

Performance measure
The suggested algorithm is executed in PYTHON (Spyder 3.8), and the simulation results are compared with some of the best methods available in the literature and are found to be significantly better.

Conclusion
In this paper, a machine learning-based image clustering technique is developed using wavelet transform. During the preprocessing stage, approximation image coefficients are denoised by a fast NL-means filter which provides a better denoising effect on random noise compared to other leading methods. A suitable soft thresholding value is automatically selected based on structural criteria of the high-frequency image details is one of the key steps in the proposed algorithm. Here SRM help in eliminating the false boundaries of segmented images and in overcoming the drawbacks of clustering algorithm by reducing the over-segmentation patterns to enhance the necessary hidden image details. Finally, both image features are fused at the wavelet projection level. However, in existing methods especially in arial image analysis, the segmented outcome is not possible to meet the desired level, due to an under-segmentation process. Thus the post-processing stage produces only a few image features while merging in level-2 decomposition. Therefore, the article suggests a new method for standard and Arial image analysis. In the segmentation process, the usage of the clustering technique efficiently segments the crucial details of the denoised image (7.1.06) and significantly overcomes the under-segmentation issues. Therefore the algorithm is more effective in detecting the image patterns as compared to some of the leading segmentation techniques and is one of the major novelty implemented in this work, which helps in the merging process to acquire essential and meaningful features of the segmented image, as shown in Figure 7, 8, and 9. However, the watershed-based hierarchical region merging (HRM) method provides 254 segments and 120 regions, and a graph-based hierarchical region merging approach achieved 350 segments and 94 regions, which were better than K-means + HRM, compact watershed + HRM, and quick shift + HRM models. The proposed SRM dependent clustering algorithm achieved 1014 segments and 29 essential object features with residual features is shown in Figure 11. Hence the suggested model is more efficient in extracting the vital details of the arial image and recognizing the object features at the wavelet restoration level. The projected image provides a 21.73 PSNR value in 17.96 elapsed time, and is compared with some of the popular segmentation techniques along with time are illustrated in Figure 11, and Table 2. Also, the visual evaluation justifies the suggested study is better than the existing methods. The utilization of wavelet transform to preserve edge features during segmentation is one of the key features of this method, and it significantly overcomes the under segmentation patterns. The entire proposed system works better in segmenting both general and arial images and it is a key advantage in various applications. Also, object features of the arial image are well recognized by the segmented image and are critical in warfare conditions is another key contribution of this technique.