Multi-path noise-suppressing network for image classification
Patent Information
- Application Number
- IN202531130837
- Authority / Receiving Office
- IN · IN
- Patent Type
- Patents
- Current Assignee / Owner
- Filing Date
- 2025-12-23
- Publication Date
- 2026-10-22
- Estimated Expiration
- 2045-12-23
AI Technical Summary
Conventional image classification systems are sensitive to noise interference, leading to reduced accuracy and reliability, especially in real-world scenarios, and existing noise suppression methods are computationally complex, incomplete, or unsuitable for real-time applications.
A multi-path noise-suppressing network that integrates noise reduction and feature extraction in a unified architecture, using an encoder, skip connection pathway, and decoder to generate denoised images, combined with a feature fusion unit that processes noisy and denoised images through parallel paths for enhanced classification.
The network achieves robust image classification with reduced computational complexity, improved accuracy, and reliability by effectively suppressing noise across varying types and intensities, suitable for real-time applications.
Smart Images

Figure 00000035_0000
Abstract
Description
TECHNICAL FIELD
[0001] The present disclosure relates to the field of image processing andclassification. More particularly, the present disclosure relates to a multi-pathnoise-suppressing network for robust image classification in noisy environments.BACKGROUND
[0002] The following description of the related art is intended to providebackground information pertaining to the field of disclosure. This section mayinclude certain aspects of the art that may be related to various features of thepresent disclosure. However, it should be appreciated that this section is used onlyto enhance the understanding of the reader with respect to the present disclosure,and not as admissions of the prior art.
[0003] Image classification has emerged as a critical component across variousdomains including medical imaging, autonomous driving, surveillance, and objectrecognition, where convolutional neural networks have demonstrated remarkablecapabilities in learning complex patterns from visual data through hierarchicalfeature extraction. However, real-world imaging scenarios frequently introducenoise interference from various sources such as sensor limitations, transmissionerrors, and environmental factors, which significantly degrades classificationperformance and reliability of conventional deep learning architectures thatremain highly sensitive to such perturbations.
[0004] Existing approaches for addressing noise interference in imageclassification typically employ filtering-based methods, machine learning-baseddenoising techniques, or deep learning-based noise removal models as separatepreprocessing stages, however, such approaches often suffer from highcomputational complexity, incomplete noise removal leading to residual artifacts,sensitivity to varying noise types and intensities, and limited adaptability acrossdiverse imaging conditions. Furthermore, conventional architectures that processnoisy images directly without adequate noise handling mechanisms exhibitreduced accuracy, while models that rely on external noise detectors or extensivepreprocessing steps introduce additional latency and resource requirements thatrender them unsuitable for real-time applications or deployment on resourceconstrained devices.
[0005] Therefore, there exists a requirement for an improved approach that caneffectively suppress noise while simultaneously performing robust imageclassification, integrate noise reduction with feature extraction in a unifiedarchitecture, handle multiple noise types without extensive preprocessing, andmaintain computational efficiency suitable for practical deployment acrossvarious imaging applications.OBJECTS OF THE PRESENT DISCLOSURE
[0006] Some of the objects of the present disclosure, which at least oneembodiment herein satisfies are as listed herein below.
[0007] An object of the present disclosure is to provide a multi-path noisesuppressing network that enables effective noise suppression in images whilepreserving essential spatial features for robust image classification across varyingnoise types and intensities.
[0008] Another object of the present disclosure is to provide an integratedarchitecture that combines noise reduction and feature extraction in a unifiedpipeline for achieving enhanced classification accuracy with reducedcomputational complexity suitable for real-time applications.
[0009] Yet another object of the present disclosure is to provide a classificationframework that improves prediction reliability by combining outputs frommultiple processing paths to enhance robustness against noise interference indiverse imaging conditions including medical imaging and surveillanceapplications.SUMMARY
[0010] This section is provided to introduce certain objects and aspects of thepresent disclosure in a simplified form that are further described below in thedetailed description. This summary is not intended to identify the key features orthe scope of the claimed subject matter.
[0011] In an aspect, the present disclosure provides a multi-path noisesuppressing network for image classification including a noise reduction unitcoupled to a memory, the noise reduction unit including an encoder configured toextract feature maps from an input image, a bottleneck layer configured to refinethe feature maps, a skip connection pathway including a Gaussian filterconfigured to filter the input image, and a decoder configured to generate adenoised image by merging decoder features with filtered features throughelement-wise multiplication. The network includes a feature fusion unit coupledto the noise reduction unit, the feature fusion unit including a first path configuredto process the input image and a second path configured to process the denoisedimage, each path including convolutional branches and a weighting unitconfigured to output fusion weights. The network includes a classification unitcoupled to the feature fusion unit, the classification unit configured to generate aclass prediction by summing a first classification output from the first path and asecond classification output from the second path.
[0012] In another aspect, the present disclosure provides a method for imageclassification including receiving an input image at a noise reduction unit,processing the input image through an encoder to generate feature maps, filteringthe input image through a Gaussian filter at a skip connection pathway, generatinga denoised image at a decoder by merging filtered features with decoder features,processing the input image and the denoised image through parallel paths at afeature fusion unit, and generating a class prediction by summing classificationoutputs from the parallel paths at a classification unit.
[0013] Various objects, features, aspects and advantages of the inventive subjectmatter will become more apparent from the following detailed description ofpreferred embodiments, along with the accompanying drawing figures in whichlike numerals represent like components.BRIEF DESCRIPTION OF DRAWINGS
[0014] The accompanying drawings are included to provide a furtherunderstanding of the present disclosure and are incorporated in and constitute apart of this specification. The drawings illustrate exemplary embodiments of thepresent disclosure and, together with the description, serve to explain theprinciples of the present disclosure. The diagrams are for illustration only, whichthus is not a limitation of the present disclosure.
[0015] FIG. 1 illustrates an exemplary block diagram of a multi-path noisesuppressing network for image classification, in accordance with an embodimentof the present disclosure.
[0016] FIG. 2 illustrates an exemplary block diagram of a Selective FeatureFusion (SFF) block, in accordance with an embodiment of the present disclosure.
[0017] FIG. 3 illustrates an exemplary schematic representation of a NoiseReduction Network (NRN), in accordance with an embodiment of the presentdisclosure.
[0018] FIG. 4 illustrates an exemplary schematic representation of a SelectiveFeature Fusion Network (SFF-Net), in accordance with an embodiment of thepresent disclosure.
[0019] FIG. 5 illustrates exemplary visual representations comparing noisy inputimages with corresponding denoised output images, pixel difference maps, andhistogram distributions for medical imaging samples including chest X-ray andbrain MRI images, in accordance with an embodiment of the present disclosure.
[0020] FIG. 6 illustrates exemplary performance evaluation results depictingnoisy input and denoised output images with corresponding Peak Signal-to-NoiseRatio (PSNR) and Structural Similarity Index Measure (SSIM) values, along withautoencoder training loss curves for chest X-ray and brain MRI datasets, inaccordance with an embodiment of the present disclosure.
[0021] FIG. 7 illustrates an exemplary flow diagram depicting a method for imageclassification using a multi-path noise-suppressing network, in accordance with anembodiment of the present disclosure.DETAILED DESCRIPTION
[0022] The ensuing description provides exemplary embodiments only, and is notintended to limit the scope, applicability, or configuration of the disclosure.Rather, the ensuing description of the exemplary embodiments will provide thoseskilled in the art with an enabling description for implementing an exemplaryembodiment. It should be understood that various changes may be made in thefunction and arrangement of elements without departing from the spirit and scopeof the disclosure as set forth.Definitions:Skip Connection Pathway: A pathway that can allow features from an input imageto bypass intermediate processing layers and merge directly with decoder features,thereby preserving spatial information during noise reduction operations.Feature Fusion: A technique that can combine feature representations extractedthrough multiple receptive field sizes using learned fusion weights to generateenriched feature maps for classification.Equal-Weight Score Fusion: A fusion strategy that can combine classificationoutputs from multiple processing paths by summation to generate a final classprediction with enhanced robustness.
[0023] An aspect of the present disclosure relates to a system for imageclassification with noise suppression including a processor, a memory coupled tothe processor, and a noise reduction unit coupled to the memory and configured toreceive an input image. The noise reduction unit can include an encoderconfigured to output feature maps, a skip connection pathway including aGaussian filter configured to output a first feature map, and a decoder configuredto merge the first feature map with decoder features through element-wisemultiplication and output a denoised image. The system can further include afeature fusion unit including a first path configured to receive the input image anda second path configured to receive the denoised image, each of the first path andthe second path including convolutional branches coupled to a weighting unit, anda classification unit configured to sum a first classification output and a secondclassification output to output a class prediction.
[0024] Various embodiments of the present disclosure can be described usingFIGs. 1 to 7.
[0025] FIG. 1 illustrates an exemplary block diagram of a multi-path noisesuppressing network for image classification, in accordance with an embodimentof the present disclosure.
[0026] Referring to FIG. 1, a system (100) for image classification with noisesuppression can include but not limited to a processor (102), a memory (104)coupled to the processor (102), a noise reduction unit (108) coupled to thememory (104), a feature fusion unit (122) coupled to the noise reduction unit(108), and a classification unit (132) coupled to the feature fusion unit (122). Thesystem (100) can establish a multi-path processing architecture by implementingparallel processing of an input image (106) and a denoised image (120), whichcan enable enhanced classification accuracy by combining predictions from bothnoisy and denoised representations.
[0027] In an embodiment, the processor (102) of the system (100) can operate byexecuting computational tasks required for noise reduction and imageclassification through processing architectures that enable concurrent handling offeature extraction, noise suppression, and classification operations. The processor(102) can function by coordinating data flow between the noise reduction unit(108), the feature fusion unit (122), and the classification unit (132),implementing task scheduling that can ensure processing operations receivetimely execution, and maintaining system state information that may enablecoordinated classification decisions.
[0028] In an embodiment, the memory (104) coupled to the processor (102) canoperate by storing instructions which, when executed by the processor (102),cause the system (100) to perform noise reduction on the input image (106),extract features through multiple processing paths, and generate a class predictionthrough fusion of classification outputs. The memory (104) can function byproviding data access for image processing operations, implementing bufferstructures for intermediate processing results, and maintaining configurationparameters that may define operational characteristics of the noise reduction unit(108) and the feature fusion unit (122).
[0029] In an embodiment, the noise reduction unit (108) coupled to the memory(104) can operate by receiving the input image (106) and generating the denoisedimage (120) through an encoder-decoder architecture with skip connections thatpreserve essential spatial features during noise suppression. The noise reductionunit (108) can include the encoder (110), a bottleneck layer (114), the skipconnection pathway (116), and the decoder (118), which can coordinate to extracthierarchical features, compress feature representations, preserve spatialinformation, and reconstruct a denoised output.
[0030] In an embodiment, the encoder (110) of the noise reduction unit (108) canoperate by receiving the input image (106) and outputting the feature maps (112)through hierarchical feature extraction that captures local and contextual patternsfrom the input image (106). The encoder (110) can include a first convolutionallayer (110-2), a batch normalization layer (110-4), a second convolutional layer(110-6), and a dropout layer (110-8), which can function by progressivelyextracting and refining feature representations while preventing overfitting duringtraining.
[0031] In an embodiment, the first convolutional layer (110-2) of the encoder(110) can operate by receiving the input image (106) and outputting a first featureset through convolution operations that capture essential local features from theinput image (106). The first convolutional layer (110-2) can implementconvolution with a kernel size in a range of 1x1 to 7x7, preferably 3x3, and afilter count in a range of 32 to 128, preferably 64, that may extract spatial patternsand edge information from the input image (106), enabling initial featurerepresentation learning with ReLU activation that introduces non-linearity.
[0032] In an embodiment, the batch normalization layer (110-4) coupled to thefirst convolutional layer (110-2) can operate by receiving the first feature set fromthe first convolutional layer (110-2) and outputting a normalized feature setthrough normalization operations that stabilize learning and accelerateconvergence. The batch normalization layer (110-4) can function by normalizingactivations across a batch, which can ensure stable gradient flow during trainingand improve generalization performance of the encoder (110).
[0033] In an embodiment, the second convolutional layer (110-6) coupled to thebatch normalization layer (110-4) can operate by receiving the normalized featureset from the batch normalization layer (110-4) and outputting a second feature setthrough convolution operations that expand feature representation capacity. Thesecond convolutional layer (110-6) can implement convolution with a kernel sizein a range of 1x1 to 7x7, preferably 3x3, and a filter count in a range of 64 to 256,preferably 128, that may capture more complex patterns and higher-levelabstractions from the normalized feature set with ReLU activation following batchnormalization.
[0034] In an embodiment, the dropout layer (110-8) coupled to the secondconvolutional layer (110-6) can operate by receiving the second feature set fromthe second convolutional layer (110-6) and outputting the feature maps (112)through regularization that randomly deactivates neurons during training. Thedropout layer (110-8) can implement a dropout rate in a range of 0.1 to 0.5,preferably 0.3, that may mitigate overfitting by preventing co-adaptation ofneurons, enhancing the generalization ability of the encoder (110).
[0035] In an embodiment, the bottleneck layer (114) coupled between the encoder(110) and the decoder (118) can operate by receiving the feature maps (112) fromthe encoder (110) and outputting compressed feature maps to the decoder (118)through feature compression that retains essential information while reducingdimensionality. The bottleneck layer (114) can implement convolution with akernel size in a range of 1x1 to 5x5, preferably 3x3, and a filter count in a rangeof 64 to 256, preferably 128, with batch normalization and ReLU activation thatmay transform high-dimensional feature maps (112) into a compact latentrepresentation, reducing redundancy while preserving critical details needed foraccurate reconstruction.
[0036] In an embodiment, the skip connection pathway (116) of the noisereduction unit (108) can operate by receiving the input image (106) and outputtingthe first feature map through a parallel processing path that preserves spatialinformation for merging with the decoder features (118-6). The skip connectionpathway (116) can include a convolution layer (116-2), a batch normalizationlayer (116-4), and the Gaussian filter (116-6), where the Gaussian filter (116-6)receives the input image (106) and outputs a filtered image, and the first featuremap is generated from the filtered image through the convolution layer (116-2)and the batch normalization layer (116-4).
[0037] In an embodiment, the Gaussian filter (116-6) of the skip connectionpathway (116) can operate by receiving the input image (106) and outputting thefiltered image through spatial filtering that applies Gaussian smoothing to reduceinitial noise while preserving structural information. The Gaussian filter (116-6)can include a kernel of size in a range of 3x3 to 7x7, preferably 5x5, and astandard deviation in a range of 0.5 to 3.0, preferably 1.5, which may provideeffective noise attenuation while maintaining edge information and spatial detailsessential for accurate image reconstruction.
[0038] In an embodiment, the convolution layer (116-2) of the skip connectionpathway (116) can operate by receiving the filtered image from the Gaussian filter(116-6) and outputting a spatial feature set through convolution operations thatextract features from the filtered image. The convolution layer (116-2) canimplement convolution with a kernel size in a range of 1x1 to 5x5, preferably3x3, and a filter count in a range of 64 to 256, preferably 128, with ReLUactivation that may capture spatial patterns from the filtered image for subsequentnormalization.
[0039] In an embodiment, the batch normalization layer (116-4) coupled to theconvolution layer (116-2) can operate by receiving the spatial feature set from theconvolution layer (116-2) and outputting the first feature map throughnormalization operations that prepare the spatial feature set for merging with thedecoder (118). The batch normalization layer (116-4) of the skip connectionpathway (116) can function by stabilizing feature distributions, which can ensureeffective combination of skip connection features with the decoder features (118-6) during a merging operation.
[0040] In an embodiment, the decoder (118) of the noise reduction unit (108) canoperate by receiving the compressed feature maps from the bottleneck layer (114)and outputting the decoder features (118-6), merging the first feature map with thedecoder features (118-6) through element-wise multiplication, and outputting thedenoised image (120). The decoder (118) can include a mask generator (118-2), amultiplication unit (118-4), and transpose convolution layers that mayprogressively upsample features while incorporating spatial information from theskip connection pathway (116).
[0041] In an embodiment, the mask generator (118-2) of the decoder (118) canoperate by receiving the first feature map from the skip connection pathway (116)and outputting a weight mask through attention processing that determinesimportance of each feature for merging. The mask generator (118-2) can include aconvolution layer with a kernel size of 1x1 and a sigmoid function, where theconvolution layer of the mask generator (118-2) processes the first feature mapand the sigmoid function scales values between 0 and 1 to control contribution ofeach feature during merging, enabling attention-modulated skip connectionfusion.
[0042] In an embodiment, the multiplication unit (118-4) of the decoder (118) canoperate by receiving the weight mask from the mask generator (118-2) and thedecoder features (118-6), performing element-wise multiplication, and outputtingmerged features (118-8). The multiplication unit (118-4) can function bymodulating the decoder features (118-6) with the weight mask, which can enableselective incorporation of skip connection information based on learned attentionweights, improving feature retention and reconstruction quality. The mergedfeatures (118-8) can be computed by element-wise multiplication of the decoderfeatures (118-6) with the weight mask, where the weight mask modulatescontribution of decoder features based on learned attention weights from the firstfeature map.
[0043] In an embodiment, the decoder (118) can generate the denoised image(120) from the merged features (118-8) through transpose convolution layers forupsampling followed by a final convolution layer with a kernel size in a range of1x1 to 5x5, preferably 3x3, a filter count of 3 corresponding to RGB channels,and a sigmoid activation that normalizes pixel values to a valid range between 0and 1. The decoder (118) can function by reconstructing spatial details from themerged features (118-8), which can produce the denoised image (120) thatpreserves essential structural information while suppressing noise artifacts.
[0044] In an embodiment, the noise reduction unit (108) can optimize noisesuppression performance by implementing a combination of Structural SimilarityIndex Measure (SSIM) Loss and Mean Squared Error (MSE) Loss that enhancesperceptual image quality during training. The SSIM Loss can operate bycomputing structural similarity between a ground truth image and the denoisedimage (120), where a loss value equals one minus an SSIM score, which canensure that the denoised image (120) retains structural integrity and perceptualquality. The MSE Loss can operate by computing pixel-wise differences betweenthe ground truth image and the denoised image (120), which can minimizereconstruction errors and improve image fidelity. A combined loss function canenable the decoder (118) to generate the denoised image (120) that preserves bothstructural features and pixel-level accuracy.
[0045] In an embodiment, the feature fusion unit (122) coupled to the noisereduction unit (108) can operate by receiving the input image (106) through thefirst path (124) and the denoised image (120) through the second path (126),processing both images through parallel feature extraction pipelines includingSFF blocks, and outputting feature maps for classification. The feature fusion unit(122) can function by implementing multi-path processing that extractscomplementary features from both noisy and denoised representations, enablingrobust classification through combination of information from both processingpaths. The parallel processing of the input image (106) through the first path (124)and the denoised image (120) through the second path (126) enables theclassification unit (132) to receive feature maps from both noisy and noisesuppressed representations, where the first classification output (132-2) capturesclassification-relevant features that may be attenuated during denoising and thesecond classification output (132-4) captures enhanced features from the noisesuppressed representation.
[0046] In an embodiment, the first path (124) of the feature fusion unit (122) canoperate by receiving the input image (106) and extracting multi-scale featuresthrough the convolutional branches (128) coupled to the weighting unit (130). Thefirst path (124) can function by processing the original noisy representation tocapture classification-relevant features that may complement features extractedfrom the denoised image (120) in the second path (126).
[0047] In an embodiment, the second path (126) of the feature fusion unit (122)can operate by receiving the denoised image (120) and extracting multi-scalefeatures through the convolutional branches (128) coupled to the weighting unit(130). The second path (126) can function by processing the noise-suppressedrepresentation to capture enhanced classification features with reduced noiseinterference.
[0048] In an embodiment, the convolutional branches (128) in each of the firstpath (124) and the second path (126) can be implemented within an SFF block,including a first branch including the Conv2D layer of kernel size 1x1 with theactivation layer implementing ReLU activation, a second branch including theConv2D layer of kernel size in a range of 2x2 to 5x5, preferably 3x3, with theactivation layer implementing ReLU activation, and a third branch including thedepthwise Conv2D layer of kernel size in a range of 3x3 to 7x7, preferably 5x5,and a dilation rate in a range of 1 to 4, preferably 2, followed by the Conv2D layerand the activation layer implementing ReLU activation. The first branch cancapture fine-grained local features with a filter count in a range of 32 to 128,preferably 64, without expanding a receptive field, the second branch can extractmid-range spatial patterns with a filter count in a range of 64 to 256, preferably128, and the third branch can capture long-range dependencies efficiently throughdilated depthwise convolution with a filter count in a range of 64 to 256,preferably 128, without increasing computational complexity. Each of the firstbranch, the second branch, and the third branch can implement ReLU activationfollowing convolution operations.
[0049] In an embodiment, the weighting unit (130) of the feature fusion unit (122)can operate by receiving outputs from the convolutional branches (128) andoutputting fusion weights and feature maps through attention-based weighting thatdetermines optimal combination of multi-scale features. The weighting unit (130)can be implemented through the attention weight generation path of the SFFblock, including the batch normalization layer (110-4), the pooling layer (130-2)implementing global average pooling, the dense layer, the activation layerimplementing softmax activation, and the reshape layer, which can function byextracting global contextual information and assigning importance scores to eachof the convolutional branches (128) for feature fusion through the attentionweighted multiplication unit and the element wise add unit.
[0050] In an embodiment, the global average pooling layer (130-2) of theweighting unit (130) can operate by receiving outputs from the convolutionalbranches (128) and outputting pooled features through spatial averaging thatcondenses feature maps into compact representations. The global average poolinglayer (130-2) can function by extracting global contextual information from eachbranch output, which can provide a basis for computing attention weights thatdetermine contribution of each receptive field size to a final featurerepresentation.
[0051] In an embodiment, the softmax layer (130-4) coupled to the global averagepooling layer (130-2) can operate by receiving the pooled features from the globalaverage pooling layer (130-2) and outputting the fusion weights through softmaxnormalization that converts the pooled features into probability distributions. Thesoftmax layer (130-4) can function by assigning the fusion weights correspondingto each of the convolutional branches (128), where the fusion weights sum tounity and indicate relative importance of each receptive field for a current input,enabling adaptive feature combination. A final fused feature map can becomputed as a weighted sum of outputs from the first branch, the second branch,and the third branch using corresponding fusion weights.
[0052] In an embodiment, the classification unit (132) coupled to the featurefusion unit (122) can operate by receiving the feature maps from the first path(124) and the second path (126), generating classification outputs from each path,and combining the classification outputs to generate the class prediction. Theclassification unit (132) can function by implementing equal-weight score fusionthat combines predictions from noisy and denoised image processing, which can enhance classification robustness by leveraging complementary information fromboth representations.
[0053] In an embodiment, the classification unit (132) can receive the featuremaps from the first path (124) and output the first classification output (132-2)through classification processing that maps extracted features to classprobabilities. The first classification output (132-2) can represent probabilitydistributions over possible classes computed from the input image (106) processedthrough the first path (124), which can capture classification information from theoriginal noisy representation. The classification processing can include a globalaverage pooling operation followed by a fully connected layer with softmaxactivation that generates the probability distribution.
[0054] In an embodiment, the classification unit (132) can receive the featuremaps from the second path (126) and output the second classification output (132-4) through classification processing that maps extracted features to classprobabilities. The second classification output (132-4) can represent probabilitydistributions over possible classes computed from the denoised image (120)processed through the second path (126), which can capture classificationinformation from the noise-suppressed representation.
[0055] In an embodiment, the classification unit (132) can include a summationunit (132-6) configured to receive the first classification output (132-2) and thesecond classification output (132-4) and output a combined output throughelement-wise addition that aggregates probability scores from both processingpaths, implementing the equal-weight score fusion strategy.
[0056] In an embodiment, the classification unit (132) can include an argmax unit(132-8) coupled to the summation unit (132-6) and configured to receive thecombined output from the summation unit (132-6) and output the class predictionthrough argmax selection that identifies a class with a highest combinedprobability score. The argmax unit (132-8) can function by determining a finalclassification result based on fused predictions, which can provide enhancedrobustness compared to single-path classification by combining evidence fromboth noisy and denoised image representations.
[0057] FIG. 2 illustrates an exemplary block diagram of a Selective FeatureFusion (SFF) block, in accordance with an embodiment of the present disclosure.
[0058] Referring to FIG. 2, an SFF block (200) can operate by receiving an inputfeature map (202) and outputting an output feature map (232) through multi-scalefeature extraction and attention-based fusion. The SFF block (200) can include anattention weight generation path (204) coupled to a batch normalization layer(110-4) and a pooling layer (130-2), a dense layer (206), an activation layerimplementing softmax activation (208), and a reshape layer (210) for generatingattention weights. The SFF block (200) can include a first convolutional branchincluding a Conv2D layer (212) and an activation layer implementing ReLUactivation (216), a second convolutional branch including a Conv2D layer (214)and an activation layer implementing ReLU activation (218), and a thirdconvolutional branch including a depthwise Conv2D layer (220), a Conv2D layer(222), and an activation layer implementing ReLU activation (224). The SFFblock (200) can include a GetItem layer (226) for selecting attention weightscorresponding to each branch, an attention weighted multiplication unit (228) formultiplying branch outputs with corresponding attention weights, and an elementwise add unit (230) for combining weighted branch outputs to generate the outputfeature map (232).
[0059] In an embodiment, each of the first path (124) and the second path (126)of the feature fusion unit (122) can include a Selective Feature Fusion (SFF) block(200) configured to receive an input feature map (202) and output an outputfeature map (232) through multi-scale feature extraction and attention-basedfusion. The SFF block (200) can include an attention weight generation path(204), convolutional branches, and a fusion mechanism that combines branchoutputs using learned attention weights to generate enriched featurerepresentations for classification.
[0060] In an embodiment, the attention weight generation path (204) of the SFFblock (200) can operate by receiving the input feature map (202) and outputtingattention weights for each convolutional branch through sequential processinglayers. The attention weight generation path (204) can include the batchnormalization layer (110-4) configured to normalize the input feature map (202), apooling layer (130-2) configured to perform global average pooling on normalizedfeatures, a dense layer (206) configured to learn channel-wise importance, anactivation layer implementing softmax activation (208) configured to generatenormalized attention weights summing to unity, and a reshape layer (210)configured to reshape the attention weights for element-wise multiplication withbranch outputs.
[0061] In an embodiment, the dense layer (206) of the attention weight generationpath (204) can operate by receiving pooled features from the pooling layer (130-2)and outputting learned feature representations through fully connected operationswith a number of output units corresponding to a number of convolutionalbranches, preferably 3. The activation layer implementing softmax activation(208) coupled to the dense layer (206) can operate by receiving learned featurerepresentations and outputting attention weights through softmax normalizationthat converts raw scores into probability distributions, where each attentionweight indicates relative importance of a corresponding convolutional branch forprocessing the input feature map (202).
[0062] In an embodiment, the SFF block (200) can include a first convolutionalbranch configured to capture fine-grained local features from the input featuremap (202). The first convolutional branch can include a Conv2D layer (212)implementing convolution with a kernel size of 1x1 and a filter count in a rangeof 32 to 128, preferably 64, and an activation layer implementing ReLU activation(216) coupled to the Conv2D layer (212), where the first convolutional branchoutputs a first branch feature map capturing local spatial patterns withoutexpanding a receptive field.
[0063] In an embodiment, the SFF block (200) can include a secondconvolutional branch configured to extract mid-range spatial patterns from theinput feature map (202). The second convolutional branch can include a Conv2Dlayer (214) implementing convolution with a kernel size in a range of 2x2 to 5x5,preferably 3x3, and a filter count in a range of 64 to 256, preferably 128, and anactivation layer implementing ReLU activation (218) coupled to the Conv2Dlayer (214), where the second convolutional branch outputs a second branchfeature map capturing intermediate-scale spatial dependencies.
[0064] In an embodiment, the SFF block (200) can include a third convolutionalbranch configured to capture long-range dependencies from the input feature map(202) through dilated depthwise convolution. The third convolutional branch caninclude a depthwise Conv2D layer (220) implementing depthwise convolutionwith a kernel size in a range of 3x3 to 7x7, preferably 5x5, and a dilation rate in arange of 1 to 4, preferably 2, a Conv2D layer (222) implementing pointwiseconvolution with a kernel size of 1x1 and a filter count in a range of 64 to 256,preferably 128, and an activation layer implementing ReLU activation (224)coupled to the Conv2D layer (222), where the third convolutional branch outputsa third branch feature map capturing long-range spatial dependencies efficientlywithout increasing computational complexity.
[0065] In an embodiment, the SFF block (200) can include a GetItem layer (226)configured to receive the attention weights from the reshape layer (210) andoutput individual attention weight values corresponding to each of the firstconvolutional branch, the second convolutional branch, and the thirdconvolutional branch. The GetItem layer (226) can function by indexing theattention weights to extract a first attention weight for the first convolutionalbranch, a second attention weight for the second convolutional branch, and a thirdattention weight for the third convolutional branch.
[0066] In an embodiment, the SFF block (200) can include an attention weightedmultiplication unit (228) configured to receive branch feature maps from the firstconvolutional branch, the second convolutional branch, and the thirdconvolutional branch, and corresponding attention weights from the GetItem layer(226), and perform element-wise multiplication to generate weighted branchfeature maps. The attention weighted multiplication unit (228) can function byscaling each branch feature map with its corresponding attention weight, wherethe first branch feature map is multiplied with the first attention weight, thesecond branch feature map is multiplied with the second attention weight, and thethird branch feature map is multiplied with the third attention weight.
[0067] In an embodiment, the SFF block (200) can include an element wise addunit (230) configured to receive weighted branch feature maps from the attentionweighted multiplication unit (228) and output the output feature map (232)through element-wise addition that aggregates multi-scale features. The elementwise add unit (230) can function by summing the weighted first branch featuremap, the weighted second branch feature map, and the weighted third branchfeature map to generate the output feature map (232) that combines local, midrange, and long-range spatial information with optimal fusion weights learnedthrough the attention weight generation path (204).
[0068] FIG. 3 illustrates an exemplary schematic representation of a NoiseReduction Network (NRN), in accordance with an embodiment of the presentdisclosure.
[0069] Referring to FIG. 3, a schematic representation (300) of the noisereduction unit (108) depicts the encoder-decoder architecture with the skipconnection pathway (116). The input image (106) can be received and processedthrough the encoder (110), the bottleneck layer (114), and the decoder (118) alonga primary processing path. Concurrently, the input image (106) can be processedthrough the skip connection pathway (116) including the Gaussian filter (116-6),where filtered features can be merged with decoder features through element-wisemultiplication at the decoder (118) to generate the denoised output.
[0070] In an embodiment, the schematic representation (300) can depict data flowwhere the input image (106) enters both the encoder (110) and the Gaussian filter(116-6) of the skip connection pathway (116) simultaneously. The encoder (110)can extract hierarchical feature maps through successive convolutional layers, thebottleneck layer (114) can compress the feature maps into a compact latentrepresentation, and the decoder (118) can upsample and reconstruct features whilereceiving preserved spatial information from the skip connection pathway (116),enabling generation of the denoised image with retained structural details.
[0071] FIG. 4 illustrates an exemplary schematic representation of a SelectiveFeature Fusion Network (SFF-Net), in accordance with an embodiment of thepresent disclosure.
[0072] Referring to FIG. 4, a schematic representation (400) of the feature fusionunit (122) depicts the processing pipeline for feature extraction and classification.The input image (106) can be received and processed through an initialconvolutional layer, a batch normalization layer (110-4), SFF blocks within thefeature fusion unit (122), and a pooling layer (130-2) to generate featurerepresentations for classification. The architecture can enable multi-scale featureextraction through the SFF blocks with attention-based fusion for generatingdiscriminative feature maps.
[0073] In an embodiment, the schematic representation (400) can depict thefeature fusion unit (122) receiving the input image (106) having dimensions of112x112x3 and progressively extracting features through a convolutional layerwith kernel size 3x3 and filter count 32, the batch normalization layer (110-4) forfeature normalization, a first SFF block (200) with filter count 64, a second SFFblock (200) with filter count 128, and the pooling layer (130-2) that condensesextracted features into a compact representation for subsequent classificationthrough a fully connected layer with softmax activation outputting n classprobabilities corresponding to a number of classification categories.
[0074] FIG. 5 illustrates exemplary visual representations comparing noisy inputimages with corresponding denoised output images, pixel difference maps, andhistogram distributions for medical imaging samples including chest X-ray andbrain MRI images, in accordance with an embodiment of the present disclosure.
[0075] Referring to FIG. 5, visual representations (500) depict comparativeresults for medical imaging samples including chest X-ray images and brain MRIimages. Each row can display a noisy image, a corresponding denoised imagegenerated by the noise reduction unit (108), a pixel difference map indicatingnoise removed, and histogram distributions depicting pixel intensity frequenciesfor both noisy and denoised images. The histogram comparisons can demonstratethat denoised images exhibit smoother and more concentrated intensitydistributions compared to noisy images with irregular distributions, validatingeffectiveness of the noise reduction unit (108) in preserving structural detailswhile suppressing noise artifacts.
[0076] In an embodiment, the visual representations (500) can demonstrate thathistogram distributions of noisy images exhibit wide and irregular intensitypatterns indicative of noise contamination, while histogram distributions ofcorresponding denoised images exhibit smooth and concentrated intensity patternsindicative of effective noise suppression with preserved structural information.The pixel difference maps can display removed noise artifacts, confirming that thenoise reduction unit (108) effectively separates noise components from essentialimage content while retaining diagnostic features for subsequent classificationthrough the feature fusion unit (122).
[0077] FIG. 6 illustrates exemplary performance evaluation results depictingnoisy input and denoised output images with corresponding Peak Signal-to-NoiseRatio (PSNR) and Structural Similarity Index Measure (SSIM) values, along withautoencoder training loss curves for chest X-ray and brain MRI datasets, inaccordance with an embodiment of the present disclosure.
[0078] Referring to FIG. 6, performance evaluation results (600) depict noisyinput images with corresponding denoised output images annotated with PeakSignal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM)values for chest X-ray and brain MRI datasets. The autoencoder training losscurves can depict train loss and validation loss over epochs, demonstrating steadyconvergence and minimal overfitting during training of the noise reduction unit(108). The PSNR values exceeding 31 dB for chest X-ray images and exceeding47 dB for brain MRI images, along with SSIM values exceeding 0.98, canvalidate effective noise suppression with preserved image quality.
[0079] In an embodiment, the performance evaluation results (600) candemonstrate that the noise reduction unit (108) achieves PSNR values in a rangeof 31 to 32 dB, preferably 31.86 dB and 31.14 dB, with SSIM values of 0.9804for chest X-ray denoising tasks, and PSNR values in a range of 47 to 49 dB,preferably 48.09 dB and 47.75 dB, with SSIM values in a range of 0.9964 to0.9966 for brain MRI denoising tasks. The higher PSNR and SSIM values forbrain MRI images can indicate superior noise suppression performance on imageswith distinct structural features compared to chest X-ray images with more diffusepatterns.
[0080] In an embodiment, the performance evaluation results (600) can includeautoencoder training loss curves depicting convergence behavior of the noisereduction unit (108) during training. The training loss curves can depict train lossand validation loss decreasing from initial values exceeding 0.5 to values below0.1 over 50 training epochs, where the train loss and validation loss curvesconverge and stabilize after approximately 20 epochs, demonstrating efficientlearning without overfitting. The convergence behavior can validate effectivenessof the combined SSIM Loss and MSE Loss function in training the encoder (110),the skip connection pathway (116), and the decoder (118) for optimal noisesuppression performance.
[0081] FIG. 7 illustrates an exemplary flow diagram depicting a method for imageclassification using a multi-path noise-suppressing network, in accordance with anembodiment of the present disclosure.
[0082] Referring to FIG. 7, a method (700) for image classification with noisesuppression can include receiving (702) the input image (106), processing (704)the input image (106) at the encoder (110), filtering (706) the input image (106)through the Gaussian filter (116-6), receiving (708) the feature maps (112) at thedecoder (118), processing (710) through parallel paths at the feature fusion unit(122), outputting (712) fusion weights at the weighting unit (130), outputting(714) classification outputs at the classification unit (132), and summing (716) theclassification outputs to generate the class prediction.
[0083] In an embodiment, block (702) receiving the input image (106) at thenoise reduction unit (108) coupled to the memory (104) can operate by acquiringthe input image (106) for noise suppression and feature extraction processing. Areceiving process can function by loading the input image (106) into the noisereduction unit (108), which can initiate parallel processing through the encoder(110) and the skip connection pathway (116) for subsequent noise reductionoperations.
[0084] In an embodiment, block (704) processing the input image (106) at theencoder (110) of the noise reduction unit (108), the encoder (110) outputting thefeature maps (112) can operate by executing hierarchical feature extractionthrough the first convolutional layer (110-2), the batch normalization layer (110-4), the second convolutional layer (110-6), and the dropout layer (110-8). Aprocessing operation of block (704) can function by progressively extractingspatial patterns and contextual information from the input image (106), generatingthe feature maps (112) that capture essential characteristics for noise reductionand classification.
[0085] In an embodiment, block (706) filtering the input image (106) through theGaussian filter (116-6) at the skip connection pathway (116) of the noisereduction unit (108), the skip connection pathway (116) outputting the first featuremap can operate by applying Gaussian smoothing with a kernel of size in a rangeof 3x3 to 7x7, preferably 5x5, and a standard deviation in a range of 0.5 to 3.0,preferably 1.5, to the input image (106) and extracting spatial features through theconvolution layer (116-2) and the batch normalization layer (116-4). A filteringprocess of block (706) can function by reducing initial noise while preservingstructural information, generating the first feature map that retains spatial detailsfor merging with the decoder features (118-6).
[0086] In an embodiment, block (708) receiving the feature maps (112) at thedecoder (118) of the noise reduction unit (108), the decoder (118) outputting thedecoder features (118-6), merging the first feature map with the decoder features(118-6) through element-wise multiplication at the decoder (118), the decoder(118) outputting the denoised image (120) can operate by reconstructing an imagefrom compressed features while incorporating spatial information from the skipconnection pathway (116). A decoding process of block (708) can function bygenerating the weight mask through the mask generator (118-2), performingelement-wise multiplication through the multiplication unit (118-4), andreconstructing the denoised image (120) from the merged features (118-8).
[0087] In an embodiment, block (710) processing the input image (106) throughthe first path (124) and the denoised image (120) through the second path (126) atthe feature fusion unit (122) coupled to the noise reduction unit (108), each of thefirst path (124) and the second path (126) including the convolutional branches(128) can operate by executing parallel multi-scale feature extraction on bothnoisy and denoised representations. A processing operation of block (710) canfunction by extracting features through the first branch with kernel size 1x1, thesecond branch with kernel size in a range of 2x2 to 5x5, preferably 3x3, and thethird branch with dilated depthwise convolution of kernel size in a range of 3x3 to7x7, preferably 5x5, and dilation rate in a range of 1 to 4, preferably 2, capturinglocal, mid-range, and long-range patterns from both processing paths.
[0088] In an embodiment, block (712) outputting the fusion weights and thefeature maps at the weighting unit (130) of the feature fusion unit (122), theweighting unit (130) receiving outputs from the convolutional branches (128) canoperate by computing attention-based weights through the global average poolinglayer (130-2) and the softmax layer (130-4). A weighting process of block (712)can function by extracting global contextual information, computing importancescores for each of the convolutional branches (128), and generating weightedfeature maps that combine multi-scale information optimally for classification.
[0089] In an embodiment, block (714) outputting the first classification output(132-2) from the feature maps of the first path (124) and the second classificationoutput (132-4) from the feature maps of the second path (126) at the classificationunit (132) coupled to the feature fusion unit (122) can operate by mapping fusedfeatures from each path to class probability distributions. A classification processof block (714) can function by generating the first classification output (132-2)representing predictions from the noisy image processing and the secondclassification output (132-4) representing predictions from the denoised imageprocessing.
[0090] In an embodiment, block (716) summing the first classification output(132-2) and the second classification output (132-4) at the classification unit(132), the classification unit (132) outputting the class prediction can operate bycombining probability scores through the summation unit (132-6) and selecting afinal class through the argmax unit (132-8). A summing process of block (716)can function by aggregating evidence from both processing paths, which canenhance classification robustness by leveraging complementary information fromnoisy and denoised representations to generate the class prediction.
[0091] In an embodiment, the system (100) can implement optimization throughan Adam optimizer with a learning rate in a range of 0.0001 to 0.001, preferably0.001 for the noise reduction unit (108) and preferably 0.0001 for the featurefusion unit (122), that enables stable convergence during training of the noisereduction unit (108) and the feature fusion unit (122). The Adam optimizer canfunction by computing adaptive learning rates for each parameter through firstand second moment estimates of gradients, which can accelerate training whilemaintaining stability across varying gradient magnitudes. The feature fusion unit(122) can implement categorical cross-entropy loss for classification training,where a loss function computes a negative log-likelihood between predicted classprobabilities and ground truth labels, enabling effective learning of discriminativefeatures through the convolutional branches (128).
[0092] In an embodiment, the system (100) can implement activation functionsincluding but not limited to ReLU activation in hidden layers of the encoder(110), the decoder (118), and the convolutional branches (128) that introduce nonlinearity and accelerate training convergence. The first convolutional layer (110-2), the second convolutional layer (110-6), and the convolution layer (116-2) caneach apply ReLU activation following convolution operations, which can enablelearning of complex non-linear patterns while mitigating vanishing gradientproblems. A sigmoid activation can be implemented in the mask generator (118-2)for generating the weight mask with values between 0 and 1, and in a final layerof the decoder (118) for normalizing pixel values of the denoised image (120). Asoftmax activation can be implemented in the softmax layer (130-4) forgenerating the fusion weights and in output layers of the classification unit (132)for generating probability distributions over classes.
[0093] In an embodiment, the system (100) can be configured for training withthe input image (106) having dimensions in a range of 64x64x3 to 224x224x3,preferably 112x112x3, corresponding to height, width, and RGB color channels,which can provide standardized input for consistent feature extraction across theencoder (110) and the feature fusion unit (122). A training process can operateover a number of epochs in a range of 20 to 100, preferably 50, with a batch sizein a range of 8 to 64, preferably 16 to 32, that balances computational efficiencywith gradient estimation accuracy. The noise reduction unit (108) can be trainedusing noisy-clean image pairs, where noisy images serve as the input image (106)and corresponding clean images serve as ground truth targets, enabling supervisedlearning of noise suppression through the encoder-decoder architecture with theskip connection pathway (116).
[0094] In an exemplary embodiment, the system (100) was implemented andvalidated using medical imaging datasets including but not limited to Brain MRIimages obtained from Figshare including tumor classification cases and Chest Xray images obtained from Kaggle including radiography images for diseaseclassification. The noise reduction unit (108) achieved a Peak Signal-to-NoiseRatio (PSNR) in a range of 40 to 55 dB, preferably 48.09 dB, and a StructuralSimilarity Index Measure (SSIM) in a range of 0.95 to 1.0, preferably 0.996, onBrain MRI denoising tasks, demonstrating superior noise suppression capabilitywhile preserving diagnostic image quality. The system (100) achievedclassification accuracy in a range of 95% to 99%, preferably 97.27%, on Chest Xray classification with two classes and classification accuracy in a range of 96% to99%, preferably 98.02%, on Brain MRI classification with three classes,outperforming conventional architectures including but not limited to VGG16 andResNet50 while utilizing significantly fewer parameters.
[0095] In an exemplary embodiment, the system (100) demonstratedcomputational efficiency with floating-point operations (FLOPs) in a range of 100billion to 150 billion, preferably 120 billion, and parameters in a range of 0.4million to 0.8 million, preferably 0.59 million, in the noise reduction unit (108)and parameters in a range of 0.5 million to 1.0 million, preferably 0.72 million, ina complete classification pipeline, achieving execution time in a range of 0.3 to0.8 seconds, preferably 0.519 seconds, per image. An implementation wasperformed on a computing platform including but not limited to a Ryzenprocessor, 32 GB of RAM, and a 4 GB NVIDIA GPU, utilizing deep learningframeworks for model development and training. The system (100) was evaluatedacross different noise types including but not limited to Gaussian noise andPoisson noise at varying intensity levels, where a multi-path architectureincluding the first path (124) processing the input image (106) and the secondpath (126) processing the denoised image (120) consistently improvedclassification robustness compared to single-path approaches, validatingeffectiveness of the equal-weight score fusion implemented through thesummation unit (132-6) and the argmax unit (132-8) of the classification unit(132).
[0096] The described system (100) presents an architecture that can enableintegration of noise suppression with image classification through multi-pathprocessing, feature fusion, and equal-weight score fusion mechanisms that mayenhance classification accuracy in noisy environments. The system (100) canprovide robust image classification by utilizing parallel processing paths thatenable complementary feature extraction from both original and denoised imageswhile maintaining computational efficiency suitable for deployment acrossvarious imaging applications.
[0097] While considerable emphasis has been placed herein on the preferredembodiments, it will be appreciated that many embodiments can be made and thatmany changes can be made in the preferred embodiments without departing fromthe principles of the disclosure. These and other changes in the preferredembodiments of the disclosure will be apparent to those skilled in the art from thedisclosure herein, whereby it is to be distinctly understood that the foregoingdescriptive matter is to be interpreted merely as illustrative of the disclosure andnot as a limitation.EXAMPLESExample 1: Noise reduction performance across different noise types
[0098] The system (100) including the noise reduction unit (108) with the encoder(110), the skip connection pathway (116) with the Gaussian filter (116-6), and thedecoder (118) was tested across different noise types and intensity levels onmedical imaging datasets. Table 1 presents experimental data showing noise type,noise level, Peak Signal-to-Noise Ratio (PSNR), Structural Similarity IndexMeasure (SSIM), and Mean Absolute Error (MAE) for X-ray and Brain MRIdenoising tasks.Table 1: Performance of the system (100) across different noise types in Xray and Brain MRI denoising.0099] The results demonstrate that the noise reduction unit (108) of the system(100) achieved effective denoising capabilities across different noise types withvarying levels of effectiveness. The system (100) performed best on Poissonnoise, maintaining high PSNR values exceeding 40 dB and SSIM valuesexceeding 0.985, indicating minimal distortion post-denoising. For Gaussiannoise, the noise reduction unit (108) effectively reduced degradation withmoderate PSNR variations observed in both X-ray and Brain MRI images.Example 2: Comparative analysis of noise reduction unit with existingdenoising models
[00100] The noise reduction unit (108) of the system (100) was compared withexisting Brain MRI denoising models to evaluate computational efficiency anddenoising quality. Table 2 presents comparative data showing floating-pointoperations (FLOPs), parameters, execution time, PSNR, and SSIM for eachmodel.Table 2: Comparative analysis of the noise reduction unit (108) with existingBrain MRI denoising models
[00101] The results demonstrate that the noise reduction unit (108) achieved thelowest FLOPs of 120 billion and parameter count of 0.59 million with the fastestexecution time of 0.519 seconds while delivering the highest PSNR of 48.09 dBand SSIM of 0.996, indicating superior noise suppression and image qualityretention compared to existing denoising approaches.Example 3: Classification accuracy comparison with existing models
[00102] The system (100) including the feature fusion unit (122) with the firstpath (124), the second path (126), the convolutional branches (128), the weightingunit (130), and the classification unit (132) with the summation unit (132-6) andthe argmax unit (132-8) was compared with existing classification models. Table3 presents comparative data showing model parameters and classificationaccuracy on Brain MRI and X-ray datasets.Table 3: Comparative analysis of the system (100) with existing classificationmodels
[00103] Comparative analysis reveals that the system (100) achievedclassification accuracy of 98.02% on Brain MRI with three classes and 97.27% onChest X-ray with two classes while utilizing only 0.72 million parameters,significantly fewer than VGG16 with 14.72 million parameters, ResNet50 with23.59 million parameters, and Attention-guided CNN with 6.63 millionparameters. The results validate effectiveness of the multi-path architectureincluding the first path (124) processing the input image (106) and the secondpath (126) processing the denoised image (120), and the equal-weight score fusionimplemented through the summation unit (132-6) for enhanced classificationrobustness with reduced computational requirements.ADVANTAGES OF THE PRESENT DISCLOSURE
[00104] The present disclosure provides a system for image classification withnoise suppression that integrates noise reduction and feature extraction within aunified pipeline through the noise reduction unit and the feature fusion unit,enabling robust classification across varying noise types and intensities whilemaintaining computational efficiency with reduced parameters suitable for realtime deployment in medical imaging, autonomous driving, and surveillanceapplications.
[00105] The present disclosure provides a multi-path architecture including thefirst path processing the input image and the second path processing the denoisedimage, where the classification unit combines the first classification output andthe second classification output through equal-weight score fusion, enablingenhanced classification robustness by leveraging complementary informationfrom both noisy and denoised representations.
Claims
1. A system (100) for image classification with noise suppression, the system (100) comprising: a processor (102); a memory (104) coupled to the processor (102), the memory (104) storing instructions, the processor (102) executing the instructions; a noise reduction unit (108) coupled to the memory (104), wherein the noise reduction unit (108) receives an input image (106), the noise reduction unit (108) comprising: an encoder (110), wherein the encoder (110) receives the input image (106) and outputs feature maps (112); a skip connection pathway (116) comprising a Gaussian filter (116- 6), wherein the Gaussian filter (116-6) receives the input image (106) and outputs a filtered image and a first feature map from the filtered image; and a decoder (118), wherein the decoder (118) receives the feature maps (112) and outputs decoder features (118-6), wherein the decoder (118) merges the first feature map with the decoder features (118-6) through elementwise multiplication and outputs a denoised image (120); a feature fusion unit (122) coupled to the noise reduction unit (108), wherein the feature fusion unit (122) receives the input image (106) through a first path (124) and the denoised image (120) through a second path (126), each of the first path (124) and the second path (126) comprising a plurality of convolutional branches (128) coupled to a weighting unit (130), wherein the weighting unit (130) outputs fusion weights and feature maps; and a classification unit (132) coupled to the feature fusion unit (122), wherein the classification unit (132) receives the feature maps from the first path (124) and outputs a first classification output (132-2), receives the feature maps from the second path (126) and outputs a second classification output (132-4), sums the first classification output (132-2) and the second classification output (132-4), and outputs a class prediction.
2. The system (100) as claimed in claim 1, the encoder (110) comprising: a first convolutional layer (110-2), wherein the first convolutional layer (110-2) receives the input image (106) and outputs a first feature set; a batch normalization layer (110-4) coupled to the first convolutional layer (110-2), wherein the batch normalization layer (110-4) receives the first feature set and outputs a normalized feature set; a second convolutional layer (110-6) coupled to the batch normalization layer (110-4), wherein the second convolutional layer (110-6) receives the normalized feature set and outputs a second feature set; and a dropout layer (110-8) coupled to the second convolutional layer (110-6), wherein the dropout layer (110-8) receives the second feature set and outputs the feature maps (112).
3. The system (100) as claimed in claim 1, the noise reduction unit (108) further comprising a bottleneck layer (114) coupled between the encoder (110) and the decoder (118), wherein the bottleneck layer (114) receives the feature maps (112) from the encoder (110) and outputs compressed feature maps to the decoder (118).
4. The system (100) as claimed in claim 1, the Gaussian filter (116-6) comprising a kernel of size 5x5 and a standard deviation of 1.5, wherein the Gaussian filter (116-6) receives the input image (106) and outputs the filtered image.
5. The system (100) as claimed in claim 1, the skip connection pathway (116) comprising: a convolution layer (116-2), wherein the convolution layer (116-2) receives the filtered image and outputs a spatial feature set; and a batch normalization layer (116-4) coupled to the convolution layer (116- 2), wherein the batch normalization layer (116-4) receives the spatial feature set and outputs the first feature map.
6. The system (100) as claimed in claim 1, the decoder (118) comprising: a mask generator (118-2) receiving the first feature map, the mask generator (118-2) comprising a convolution layer and a sigmoid function, wherein the mask generator (118-2) outputs a weight mask; and a multiplication unit (118-4), wherein the multiplication unit (118-4) receives the weight mask and the decoder features (118-6) and performs element-wise multiplication and outputs merged features (118-8).
7. The system (100) as claimed in claim 1, the plurality of convolutional branches (128) comprising: a first branch comprising a convolution layer of kernel size 1x1; a second branch comprising a convolution layer of kernel size 3x3; and a third branch comprising a depthwise convolution layer of kernel size 5x5 and a dilation rate of 2.
8. The system (100) as claimed in claim 1, the weighting unit (130) comprising: a global average pooling layer (130-2), wherein the global average pooling layer (130-2) receives outputs from the plurality of convolutional branches (128) and outputs pooled features; and a softmax layer (130-4) coupled to the global average pooling layer (130- 2), wherein the softmax layer (130-4) receives the pooled features and outputs the fusion weights, wherein the fusion weights correspond to each of the plurality of convolutional branches (128).
9. The system (100) as claimed in claim 1, the classification unit (132) comprising: a summation unit (132-6), wherein the summation unit (132-6) receives the first classification output (132-2) and the second classification output (132-4) and outputs a combined output; and an argmax unit (132-8) coupled to the summation unit (132-6), wherein the argmax unit (132-8) receives the combined output and outputs the class prediction.
10. A method (700) for image classification with noise suppression, the method comprising: receiving (702) an input image (106) at a noise reduction unit (108) coupled to a memory (104); processing (704) the input image (106) at an encoder (110) of the noise reduction unit (108), the encoder (110) outputting feature maps (112); filtering (706) the input image (106) through a Gaussian filter (116-6) at a skip connection pathway (116) of the noise reduction unit (108), the skip connection pathway (116) outputting a first feature map; receiving (708) the feature maps (112) at a decoder (118) of the noise reduction unit (108), the decoder (118) outputting decoder features (118-6), merging the first feature map with the decoder features (118-6) through elementwise multiplication at the decoder (118), the decoder (118) outputting a denoised image (120); processing (710) the input image (106) through a first path (124) and the denoised image (120) through a second path (126) at a feature fusion unit (122) coupled to the noise reduction unit (108), each of the first path (124) and the second path (126) comprising a plurality of convolutional branches (128); outputting (712) fusion weights and feature maps at a weighting unit (130) of the feature fusion unit (122), the weighting unit (130) receiving outputs from the plurality of convolutional branches (128); outputting (714) a first classification output (132-2) from the feature maps of the first path (124) and a second classification output (132-4) from the feature maps of the second path (126) at a classification unit (132) coupled to the feature fusion unit (122); and summing (716) the first classification output (132-2) and the second classification output (132-4) at the classification unit (132), the classification unit (132) outputting a class prediction.
