Inter national J our nal of Inf ormatics and Communication T echnology (IJ-ICT) V ol. 15, No. 2, June 2026, pp. 508 ∼ 522 ISSN: 2252-8776, DOI: 10.11591/ijict.v15i2.pp508-522 ❒ 508 A r ob ust outlier detection based ltering f or noise r emo v al in grayscale images Ali Salem Al Rawash 1 , F arah Aini Abdullah 1 , Ahmad Kadri J unoh 2 , Abdallah Alshbeel 1 , Mohammed Banikhalid 1 1 School of Mathematical Sciences, Uni v ersity Sains Malaysia Gelugor , Penang, Malaysia 2 F aculty of Intelligent Computing, Uni v ersiti Malaysia Perlis, Arau, Malaysia Article Inf o Article history: Recei v ed Jul 11, 2025 Re vised Mar 4, 2026 Accepted Mar 30, 2026 K eyw ords: Image denoising Image restoration Interquartile range Outlier detection Salt and pepper noise ABSTRA CT Salt-and-pepper noise se v erely de grades the visual quality of digital images, par - ticularly at high noise densities where con v entional denoising techniques often f ail. Median- and mean-based lters tend to o v ersmooth images and blur ne structures when the majority of pix els within a local windo w are corrupted. This paper proposes a rob ust dual-layer denoising frame w ork for grayscale images that inte grates rank-based prescreening, interquartile range (IQR)-based statis- tical outlier detection using T uk e y fences, and a lightweight post-processing sharpening stage. In the rst layer , a rank-4 trimmed estimator suppresses e x- treme i mpulse v alues and stabilizes local statistics. In the second layer , adap- ti v e IQR thresholds are emplo yed to detect and replace residual outliers, e v en in hea vily corrupted neighborhoods. A nal step in v olving selecti v e sharpen- ing combined with mild smoothing enhanc es edge details without amplifying residual noise. Extensi v e e xperiments on standard grayscale images (Lenna, Barbara, lak e, boat, and li ving room) across s alt-and-pepper noise le v els from 10% to 90% demonstrate that the proposed approach consistently outperforms con v entional methods, including mean, median, Gaussian, modied decision based unsymmetrical trimmed median l ter (MDB UTMF), and pix el density- based lter (BPDF). Quantitati v e e v aluation indicates peak signal-to-noise ratio (PSNR) v alues reaching 38.23dB, structural similarity inde x (SSIM) v alues up to 0.99, and signicant reductions in mean squared error (MSE) , particularly at higher noise densiti es. These results conrm that the proposed frame w ork ef- fecti v ely suppresses noise while preserving edges and te xtures, making it well- suited for practical applications such as medical imaging, remote sensing, and surv eillance. This is an open access article under the CC BY -SA license . Corresponding A uthor: Ahmad Kadri Junoh F aculty of Intelligent Computing, Uni v ersity Malaysia Perlis Arau, Malaysia Email: kadri@unimap.edu.my 1. INTR ODUCTION Noise reduction remains a fundamental problem in digital image processing despite decades of re- search [1]. Digital images are representations of real-w orld scenes captured through digital sensors and stored as pix el matrices [2]. Each pix el encodes intensity information, and damage to these pix els-s uch as through noise contamination-can de grade image quality and hinder analysis. Image processing techniques aim to re- store or enhance image quality , often through ltering methods that suppress noise while preserving important J ournal homepage: http://ijict.iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Inf & Commun T echnol ISSN: 2252-8776 ❒ 509 structural features [3], [4]. Ef fecti v e noise remo v al is crucial in image processing, b ut it must be achie v ed without compromising essential image information. Satpath y et al . [5] Accurate restoration of corrupted im- age intensities while preserving structural details remains a signicant challenge. Numerous techniques ha v e been emplo yed for image denoising [6]. Salt-and-pepper noise, typically resulting from abrupt disturbances such as sensor f aults or transmission errors, leads to signicant de gradation in image quality . Although notable progress has been made in mitig ating salt and pepper noise, challenges lik e image blurring and structural dis- tortion continue to hinder performance. Thus, there remains a pressing need to impro v e the ef fecti v eness and precision of denoising techniques [7]. A major challenge in image processing is mitig ating high-density salt-and-pepper noise, a type of impulsi v e corruption in which af fected pix els assume either minimum (0) or maximum (255) intensity v alues [8]. This noise disrupts local image stati stics, causing structural distortions and loss of ne details. Critical applications-including medical imaging, remote sensing, and surv eillance-require rob ust denoising techniques that can restore visual clarity while preserving structural inte grity [9], [10]. Numerous denoising techniques ha v e been proposed. Median-based lters such as the median l - ter (MF) and its v ariants (weighted median, adapti v e median) that performing well at lo w-to-moderate noise densities by preserving edges while remo ving outliers [11], [12]. Ho we v er , at higher noise le v els (typically abo v e 50%, these methods suf fer from e xcessi v e smoothing, misclassication of v alid pix els, and loss of ne details [13]. Adv anced denoising methods aim to r emo v e noise while preserving te xtures and edges. Some approaches utilize partial dif ferential equations (PDEs), such as anisotropic second- and fourth-order dif fusion models based on con v olutional virtual electric elds and gradient v ector con v olution [14]. Additionally , the median lter remains rele v ant in modern nois e remo v al techniques, often in modied forms or combined with other methods. Examples include the tropical MF (TMF), dif ferent applied MF (D AMF), based pix el density lter (BPDF), modied decision-based unsymmetric trimmed median lter , adapti v e fuzzy switching median lter , and tropical adapti v e lter [15]. Se v eral studies ha v e attempted to address these limitations through adapti v e or weighted strat e gi es, such as the tropical median lter and decision-based lters, pro vide incremental impro v ements; ho we v er , their performance de grades signicantly when noise density e xceeds 60% [16]. Digitally acquired images are fre- quently de graded by noise, which not only reduces visual quality b ut also inuences human visual perception. In response, PDE-based models ha v e emer ged in recent years a s ef fecti v e denoising tools, of fering the k e y adv antage of reducing noise while preserving essential image details [17]. Con v entional outlier detection tech- niques construct a model of the data and classify points that de viate mark edly as outliers. Ho we v er , when the dataset contains a high proportion of outliers, these anomalous point s can distort the model itself, reducing its reliability and causing most standard detectors to f ail [18]. Despite these adv ancements, a g ap remains: e xisting techniques either e xcessi v ely smooth the i mage, f ail at high noise densities, or require impractical computational resources. This study aims to de v elop a rob ust, ef cient denoising method for grayscale images with high-density salt-and-pepper noise. The proposed tech- nique inte grates adapti v e rank ltering, statistical outlier detection (using the interquartile range), and image sharpening to achie v e superior noise suppression while preserving structural details and ne te xtures. A h ybrid denoising method is proposed to combine ra n k ltering and interquartile range (I QR)-based outlier detection tailored for high-density salt-and-pepper noise in grayscale images. Limitations of e xisting median-based and PDE-based lters are critically analyzed to sho w ho w proposed method o v ercomes these challenges. An image sharpening stage that restores ne details without amplifying residual noi se are intro- duced. Then, the method across standard test images and demonstrate superior performance o v er traditional techniques in terms of peak signal-to-noise ratio (PSNR), structural similarity inde x (SSIM), and mean squared erro (MSE) are v alidated. Although rank-based, switching, and quartile-dri v en lters ha v e been indi vidually e xplored, no e xisting denoising approach inte grates, - Rank-4 trimmed prescreening - Adapti v e T uk e y-fence outlier detection (IQR-based) - Sharpening guided by statistically cleaned estimates The no v elty of the proposed method l ies in the syner gistic combination of these three elements. Rank- 4 reduces e xtreme impulses, IQR-based statistics enable adapti v e outlier detection e v en in hea vily corrupted windo ws, and sharpening restores details suppressed by median operations. This h ybrid design produces e x- ceptional rob ustness, sustaining performance up to 80–90% nois e density , f ar be yond the limits of traditional lters. A r ob ust outlier detection based ltering for noise r emo val in gr ayscale ima g es (Ali Salem Al Rawash) Evaluation Warning : The document was created with Spire.PDF for Python.
510 ❒ ISSN: 2252-8776 2. RELA TED W ORKS Ov er the past decades, numerous image restoration techniques ha v e been de v eloped to address sal t- and-pepper noise contamination. Despite these adv ances, man y e xisting approaches demonstrate notable per - formance de gradation at ele v ated noise densities, particularly when the corruption le v el e xceeds approximately 60–70%, where maintaining structural inte grity and ne details becomes increasingly challenging [19]. Among classical solutions, median-based ltering techniques remain widely adopted due to their simplicity and ef fec- ti v eness ag ainst impulsi v e noise. The standard MF replaces each pix el with the median v alue of its surrounding neighborhood, which enables ef fecti v e suppression of isolated outliers at lo w to modera te noise le v els. Ho w- e v er , when the noise density surpasses roughly 50%, MF e xhibits e xcessi v e smoothing beha vior , leading to edge blurring and signicant loss of high-frequenc y image details [20]. T o o v ercome these limitations, se v eral enhanced median-based approaches ha v e been proposed. One such method is the Adapti v e W iener Filter , which incorporates tw o-dimensional maximum Shannon entrop y to impro v e denoising stability under high-density noise conditions [21]. Although this approach demonstrates impro v ed s uppression of impulsi v e noise, its ef fecti v eness remains limited in images with com ple x te xtures and structural v ariations, indicating the need for furt her renement. In another study , a real-time denoising frame w ork emplo ying double-nested ltering combined with morphological dilation w as introduced to remo v e high-density salt-and-pepper noise in medical imaging applications [22]. Iterati v e ltering strate gies ha v e also been e xplored, such as the iterati v e MF (IMF), which e stimates corrupted pix els using the mean intensity of detected noise-free pi x el s within a x ed windo w [23]. Although IMF performs satisf actorily at lo w noise densi- ties, its reliability deteriorates as noise density increases, since the number of v alid pix els a v ailable for accurate estimation rapidly diminishes. Other notable techniques include the adapti v e MF (AMF), which dynamically adjusts the windo w size according to local noise conditions [24]. While AMF impro v es upon the con v ent ional MF , it may inadv ertently modify uncorrupted pix els, resulting in de gradation of ne image features and edge structures. At noise densities e xceeding 60%, AMF increasingly misclassies clean pix els as noisy , leading to structural distortion [25]. Se v eral v ariants, such as the weighted MF (WMF) and the progressi v e switching MF (PSMF), ha v e been proposed to reduce unnecessary smoothing and enhance detail preserv ation. Ne v ertheless, these methods continue to f ace signicant challenges under e xtreme noise conditions, as their noise detection accurac y declines substantially at high corruption le v els [26]. Be yond median-based techniques, PDE-based methods, including anisotropic dif fusion and total v ari- ation models, ha v e been applied to salt-and-pepper noise remo v al. These approaches aim to preserv e edge information while smoothing noise through iterati v e dif fusion processes. Although ef fecti v e in maintaining structural continuity , PDE-based methods often suf fer from blurring artif acts at high noise densities and require careful parameter tuning, which limits their suitability for real-time or computationally constrained en viron- ments [27]. Despite the de v elopment of these di v erse ltering strate gies, a persistent challenge remains: balancing noise remo v al with the preserv ation of ne image details in the presence of high-density salt-and-pepper noise. The proposed method in this study addresses this g ap by combining rank ltering, statistical outlier detection using interquartil e range analysis, and image sharpening. This h ybrid approach is designed to pro vide superior noise suppression while maintaining structural inte grity , particularly in grayscale images af fected by e xtreme noise conditions. Existing lters de grade se v erely when the number of corrupted pix els e xceeds the number of clean pix els in the neighborhood. The y rely on x ed rules or trimmed medians rather than adapti v e statistical bound- aries. The proposed method lls this g ap by incorporating rank statistics + IQR statisti cs, enabling reliable outlier detection e v en in hea vily corrupted windo ws. 3. PR OPOSED METHOD This section introduces the proposed dual-layer denoising frame w ork designed to remo v e high-density salt-and-pepper noise from grayscale images. The Algorithm consists of three main components: (i) rank- based prescreening, (ii) IQR-based statistical outlier detection, and (iii) sharpening with mild smoothing. All operations are applied locally on 3 × 3 neighborhoods. Comparison of k e y features between e xisting lters and the proposed method in T able 1. Int J Inf & Commun T echnol, V ol. 15, No. 2, June 2026: 508–522 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Inf & Commun T echnol ISSN: 2252-8776 ❒ 511 3.1. Local windo w and notation F or each pix el I(x,y), a x ed 3×3 local windo w is e xtracted, W ( x, y ) = { I ( i, j ) | i ∈ [ x − 1 , x + 1] , j ∈ [ y − 1 , y + 1] } The nine v alues in the windo w are sorted in ascending order , S = { p 1 ≤ p 2 ≤ p 3 ≤ p 4 ≤ p 5 ≤ p 6 ≤ p 7 ≤ p 8 ≤ p 9 } T able 1. Comparison of k e y features between e xisting lters and the proposed method Method Rank lt. IQR stat. Switch logic Noise > 70% Edge pres. Notes Median lter No No No No No F ails when most pix els in the windo w are noisy . AMF No No Y es ∼ 60% No V ariable windo w; unstable at v ery high noise le v els. BPDF No No Y es ∼ 70% No Performance de grades when clean pix els are scarce. MDB UTMF No No Y es ∼ 70% No T rims e xtreme v alues; still dependent on x ed decision rules. Proposed Y es Y es Y es Y es Y es Dual-layer h ybrid method with statistical adapti vity . 3.2. Lay er 1: Rank-based pr escr eening (Rank = 4) Salt-and-pepper noise introduces e xtreme v alues that distort the local statistics. T o reduce their inu- ence, a rank-4 v alue is used as a trimmed estimator , R = p 4 Let the center pix el be w c = I ( x, y ) . The prescreening rule is, I 1 ( x, y ) = R , if is inconsistent w ith R w c other w ise. Where, - I 1 ( x, y ) : output of the prescreeni ng layer - R = p 4 : rank-4 rob ust estimate - w c : center pix el Rank-4 ltering stabilizes the local distrib ution and reduces strong impuls es before the statistical detection stage. 3.3. Lay er 2: IQR-based statistical outlier detection (1.5 × IQR) The second layer performs adapti v e outlier detect ion using quartile statistics deri v ed from the s orted windo w S. - Quartiles, Q 1 = p 3 , Q 3 = p 7 - Interquartile Range, I QR = Q 3 − Q 1 (1) - T uk e y Fences ( 1 . 5 × I QR ), LO L = Q 1 − 1 . 5 I QR (2) U O L = Q 3 + 1 . 5 I QR (3) A r ob ust outlier detection based ltering for noise r emo val in gr ayscale ima g es (Ali Salem Al Rawash) Evaluation Warning : The document was created with Spire.PDF for Python.
512 ❒ ISSN: 2252-8776 A v alue v in the windo w is declared an outlier if, v < LO L or v > U O L All detected outliers are replaced with the median, v ← p 5 The corrected center pix el after this layer is written as, I 2 ( x, y ) = p 5 I nal = G σ ( I sharp ) , σ ≈ 0 . 5 W ( x, y ) = { I ( i, j ) | i ∈ [ x − 1 , x + 1] , j ∈ [ y − 1 , y + 1] } Where, - Q 1 , Q 3 : rst and third quartiles - IQR: interquartile range. - LOL, UOL: lo wer and upper T uk e y limits. - p 5 : original median. - p 5 ∗ : corrected median after outlier replacement. - I 2 ( x, y ) : IQR-correc ted output pix el. The median p 5 appears in tw o stages, - In Layer 1, p 5 is implicitly related to p 4 , since rank-4 lies close to the median and helps stabilize the distri- b ution. - In Layer 2, p 5 is e xplicitly used to replace statistical outliers because it is the most rob ust estimate ag ainst impulse noise. Thus, the median supports stability in Layer 1 and rob ust replacement in Layer 2, each serving a dif ferent function. 3.4. P ost-denoising shar pening and smoothing Median-based operations may slightly smooth edges. T o restore edge sharpness, a linear unsharp- masking operator is applied, Where, I 1 : prescreened image I 2 : IQR-corrected image 1 . 5 : enhances high-frequenc y edge components − 0 . 5 : subtracts part of the smoothed component − 0 . 5 : intensity shift maintaining numerical stability A mild Gaussian blur is then applied: I nal = G σ ( I sharp ) , σ ≈ 0 . 5 where G σ denotes con v olution with a Gaussian k ernel. Rationale for the chosen weights, - 1.5 amplies edges without producing ringing artif acts (v alues abo v e 2 cause o v ershoot). - 1.5 amplies edges without producing ringing artif acts (v alues abo v e 2 cause o v ershoot). - –0.5 (bias) ensures the weighted sum remains within the v alid intensity range - σ ≈ 0 . 5 pro vides gentle smoothing that reduces sharpening noise. 3.5. P erf ormance metrics The performance of the proposed method is assessed using three standard image quality m etrics: PSNR, MSE, and SSIM. PSNR quanties the delity of the reconstructed image relati v e to the reference image, where higher v alues indicate superior reconstruction quality , and is dened as, P S N R = 10 l og 10 ( M AX 2 M S E ) (4) Int J Inf & Commun T echnol, V ol. 15, No. 2, June 2026: 508–522 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Inf & Commun T echnol ISSN: 2252-8776 ❒ 513 Where, MAX represents the maximum possibl e pix el v alue (255 for 8-bit images), and MSE is the mean squared error . MSE quanti es the a v erage squared dif ference between the original and processed images, M S E = 1 n X n i =1 ( Y i − b Y i ) 2 (5) The MSE is calculated by taking the sum of the squared dif ferences between the observ ed v alues (Y i) and the predicted v alues (Y), di vided by the number of data points (n). The SSIM formula is computed as follo ws, S S I M = (2 µ x µ y + C 1 ) + (2 σ xy + C 2 ) ( µ 2 x + µ 2 y + C 1 ) + ( σ 2 x + σ 2 y + C 2 ) (6) µ x and µ y represent the a v erage intensities (standard de viation) in the x and y directions, respecti v ely . The SSIM v alue is k ept at one using constants C 1 and C 2 . These metrics are computed for each iteration, and the results are recorded for further analysis. The nal output is e xpected to sho w signicant impro v ement in image quality with reduced noise and preserv ed detail. 3.6. Pseudocode The Algorithm 1 consists of three main components: (i) rank- based prescreening, (ii) IQR-based s ta- tistical outlier detection, and (iii) sharpening with mild smoothing. Algorithm 1: Proposed h ybrid denoising algorithm Input : Noisy grayscale image I Output: Final denoised image I f inal 1 f or eac h pi xel ( x, y ) do 2 Extract W = 3 × 3 windo w 3 Sort W → { p 1 ≤ p 2 ≤ · · · ≤ p 9 } // Layer 1: Rank-based prescreening 4 R ← p 4 5 if center pixel de viates fr om R then 6 I 1 ( x, y ) ← R 7 else 8 I 1 ( x, y ) ← center pix el // Layer 2: IQR-based outlier detection 9 Recompute sorted windo w if needed 10 Q 1 ← p 3 , Q 3 ← p 7 11 I QR ← Q 3 − Q 1 12 LO L ← Q 1 − 1 . 5 × I QR 13 H O L ← Q 3 + 1 . 5 × I QR 14 f or each value v in W do 15 if v < LO L or v > H O L then 16 v ← p 5 17 I 2 ( x, y ) ← correcte d center v alue // Sharpening and Smoothing 18 I shar p ← 1 . 5 × I 1 − 0 . 5 × I 2 − 0 . 5 19 I f in al ← GaussianBlur ( I shar p , 3 × 3 , σ ≈ 0 . 5) 3.7. Flo wchart of the pr oposed method Flo wchart illustrating the proposed dual-layer denoising frame w ork, including rank-4 prescreening, IQR-based outlier remo v al ( 1 . 5 × I QR T uk e y fences) [28]-[30], and the sharpening–smoothing post-processing stage and it sho wn in Figure 1. A r ob ust outlier detection based ltering for noise r emo val in gr ayscale ima g es (Ali Salem Al Rawash) Evaluation Warning : The document was created with Spire.PDF for Python.
514 ❒ ISSN: 2252-8776 Figure 1. Flo wchart for the method 4. RESUL T AND DISCUSSION 4.1. Experimental setup All e xperiments were implemented in Python using NumPy , SciPy , Pillo w , and scikit-image l ibraries. The proposed method w as applied on a x ed 3 × 3 windo w , using replicate padding for boundary handling. The rank-4 prescreening and 1 . 5 × I QR T uk e y fences remained constant for all test images. Synthetic salt- and-pepper noise w as generated at densities ranging from 10% to 90%, where corrupted pix els were randomly assigned v alues 0 (pepper) or 255 (salt). F or each noise le v el, v e independent noise realizations were gener - ated, and the a v erage PSNR, MSE, and SSIM v alues were reported. Experiments were performed on standard grayscale test images such as Lena, Barbara, lak e, boat and li ving room. The Figures 2 used in this study include: (a) Lenna, (b) Barbara, (c) lak e, (d) boat, and (e) li ving room sho wn in Figures 2(a)-(e). (a) (b) (c) (d) (e) Figure 2. Fi v e sample images: (a) Lenna, (b) Barbara, (c) lak e, (d) boat, and (e) li ving room Int J Inf & Commun T echnol, V ol. 15, No. 2, June 2026: 508–522 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Inf & Commun T echnol ISSN: 2252-8776 ❒ 515 4.2. Quantitati v e perf ormance The qualitati v e analysis of the proposed technique is presented here. Figure 3 illustrates the perfor - mance of t h e proposed ltering approach on the Lenna image. The upper images depict noisy v ersions of the image with salt-and-pepper noise le v els ranging from 10(100) to 30(100), representing dif ferent le v els of noise corruption. The lo wer images sho wcase the results after applying the proposed denoising method. As the noise density increases, con v entional lters often struggle to retai n structural details, leading to blurring and loss of ne te xtures. Ho we v er , the proposed method signicantly impro v es image clarit y while preserving edges and ne details, as e vident in the lo wer images. This observ ation is quantitati v ely supported by PSNR and SSIM measurements, which demonstrate superior noise remo v al and structural preserv ation compared to tra- ditional ltering techniques. The method ef fecti v ely suppresses se v ere salt-and-pepper noise while preserving k e y image features, supporting its use in optical image enhancement. 10% 20% 30% Figure 3. The results of the proposed ltering technique for Lenna images Figure 4 illustrates, as the noise density increases, the ef fecti v eness of traditional ltering met hods diminishes, leading to more visible noise artif acts. In contrast, the proposed method maintains image quality with signicantly fe wer distortions. 40% 50% 60% Figure 4. The results of the proposed ltering technique for Barbara images A r ob ust outlier detection based ltering for noise r emo val in gr ayscale ima g es (Ali Salem Al Rawash) Evaluation Warning : The document was created with Spire.PDF for Python.
516 ❒ ISSN: 2252-8776 The l tering outcomes for the Lak e image are sho wn in Figure 5, where noisy images with salt-and- pepper noise le v els ranging from 70% to 90% were processed. The proposed method signicantly reduced noise, leading to impro v ed edge denition and reduced distortion artif acts. In high noise conditions (abo v e 80%), traditional median-based lters f ail to accurately reconstruct edges, often resulting in o v er -smoothing. Ho we v er , the proposed method demonstrated stronger edge retention, as conrmed by SSIM and MSR v alues, which indicated that the ltered images had higher structural similarity and lo wer residual error compared to median ltering. These impro v ements highlight the ability of the proposed method to suppress high-density noise while preserving ne structural details, making it more ef fecti v e than con v entional denoising approaches. 70% 80% 90% Figure 5. The results of the proposed ltering technique for lak e images Quantitati v e analysis of the image ltering results is summarized in T able 2. sho wing impro v ements across all metrics. Among them, SSIM of fers a more perceptually meaningful assessment compared to MSE and PSNR, as the latter quantify absolute errors while SSIM accounts for structural and salienc y-based dif fer - ences. As noise density ris es, output image quality decreases, as illustrated in Figure 6, which compares MSE v alues for noise le v els of 10%, 20%, and 30%. The x-axis denotes noise density across v e test images, while the y-axis sho ws the corresponding MSE. T able 2. Image quality metrics results of the proposed method Images Noise MSE SSIM PSNR 0.1 8.89 37.55 0.9950 Lenna 0.2 9.10 37.44 0.9948 0.3 9.36 38.41 0.9950 0.1 10.05 37.24 0.9936 Barbara 0.2 10.19 37.18 0.9933 0.3 10.36 37.97 0.9935 0.1 9.17 38.50 0.9951 Lak e 0.2 9.66 38.27 0.9949 0.3 9.75 38.23 0.9947 0.1 7.76 38.95 0.9945 Boat 0.2 7.85 38.90 0.9943 0.3 8.17 39.00 0.9940 0.1 8.26 38.95 0.9939 Li ving Room 0.2 8.38 38.89 0.9936 0.3 8.75 38.70 0.9933 Int J Inf & Commun T echnol, V ol. 15, No. 2, June 2026: 508–522 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Inf & Commun T echnol ISSN: 2252-8776 ❒ 517 Figure 6. Graph of the MSE The proposed method’ s performance, assessed using MSE, demonstrates superior results relati v e to con v entional ltering approaches o v er v arious noise densiti es, with consistentl y lo wer MSE v alues. Although MSE gradually increases with higher noise le v els , the method ef fecti v ely preserv es image quality e v en under se v ere corruption, as indicated in T able 3 and Figure 7 illustrates the MSE v ariation across dif ferent noise densities. The x-axis represents noise density , while the y-axis represents MSE v alues. The gure highlights ho w the proposed method consistently outperforms traditional lters in reducing mean squared error . T able 3. A comparison of MSE a v erage v alues Noise 10% 20% 30% 40% 50% Mean Filter 28.12 51.23 74.23 94.32 109.21 Median Filter 28.52 52.33 78.22 98.23 112.23 Gaussian Filter 26.11 52.10 75.23 94.09 108.12 Proposed Filter 8.826 9.10 10.36 10.48 11.12 Figure 7. Graph of a comparison of MSE with dif ferent techniques A r ob ust outlier detection based ltering for noise r emo val in gr ayscale ima g es (Ali Salem Al Rawash) Evaluation Warning : The document was created with Spire.PDF for Python.