Image denoising aims to recover clean images from noisy observations while preserving fine structural details and texture information. Existing convolutional neural network (CNN)-based methods effectively capture local image structures but have limited capability for modeling long-range contextual dependencies, whereas Vision Transformer (ViT)-based approaches excel at global context modeling but often lack strong spatial inductive bias for fine texture preservation. To address these complementary limitations, we propose Convformer, a hybrid CNN–Transformer architecture for both synthetic additive white Gaussian noise (AWGN) and real-world image denoising. The proposed framework follows a residual learning paradigm and consists of a multi-domain Shallow Feature Extractor (SFE) and a Transformer-based Deep Feature Extractor (DFE). The SFE jointly exploits three complementary representations: the original spatial image, an intensity-inverted (negative) representation that enhances complementary structural and contrast information, and a frequency-domain representation obtained using DCT/IDCT transformations. These complementary features are extracted using the proposed Multi-Scale Residual Attention (MSRA) module and subsequently refined by the DFE. The DFE employs an encoder–decoder Transformer architecture equipped with the proposed Convolution-enhanced Window Multi-Head Self-Attention (Conv-WMHSA) and Convolution-enhanced Feed-Forward Network (Conv-FFN), which effectively combine long-range dependency modeling with local texture preservation through convolutional inductive bias. The network predicts the residual noise, which is subtracted from the noisy input to obtain the restored image. Extensive experiments on grayscale, color, and real-world denoising benchmarks demonstrate that Convformer achieves superior PSNR and SSIM with visually faithful reconstructions compared with representative methods. Ablation studies further verify the effectiveness of multi-domain feature extraction and CNN–Transformer integration.

Image denoising is a fundamental low-level vision task essential for restoring perceptual quality and improving downstream visual performance. Despite significant progress in deep learning-based approaches, existing methods often fail to balance noise suppression and structural detail preservation under complex noise distributions, highlighting the need for better exploitation of spatial feature dependencies within images. To tackle these issues, this article proposes SPCANet, a novel denoising framework incorporating gradient information through three specialized attention mechanisms: Pixel Attention module, Channel Attention Module, and Spatial Feature Attention Block. SPCANet effectively utilizes gradients at multiple feature processing stages to enhance image features and reduce noise. Its comprehensive approach simultaneously exploits pixel-level, channel-wise, and spatial feature dependencies, enabling robust noise reduction across various image structures. Extensive experiments are conducted on standard benchmark datasets - BSD68/CBSD68, SunHays80, and Set5/Set12 - using Gaussian noise levels of 30, 40, 50 and 60. Quantitative comparisons with multiple state-of-the-art methods, including DnCNN, MemNet, and SwinIR, demonstrate the superior performance of SPCANet, achieving an average PSNR gain of 0.93 dB and SSIM improvement of 0.111 across all datasets. Qualitative results further show that SPCANet preserves fine structures and edges more effectively than competing methods. These results confirm the robustness and generalization capability of SPCANet, providing a comprehensive and efficient solution to the general image denoising problem.

Image denoising is essential in applications such as medical imaging, remote sensing, and photography. Despite advancements in deep learning, denoising models still face key limitations. Most state-of-the-art methods increase network depth to boost performance, leading to higher computational costs, complex training, and diminishing returns. Moreover, the role of gradient information and negative image features in denoising is often overlooked, limiting the ability to capture fine structures. Our observations reveal that excessively deep networks can reduce denoising performance by introducing redundancy and complicating feature extraction. To address this, we propose MGMSDNet, a Gradient-Guided Convolutional Neural Network (CNN) with attention mechanisms that balance denoising performance and computational efficiency. MGMSDNet introduces a unique attention framework that utilizes multidirectional gradients and negative image features separately, enhancing structural preservation and noise suppression. To the best of our knowledge, this study is the first to explore multidirectional gradients for image denoising literature. MGMSDNet surpasses state-of-the-art methods on benchmark datasets, confirmed by quantitative metrics and visual comparisons. Ablation studies highlight the effectiveness of individual network components. For more details and implementation, visit our GitHub repository: MGMSDNet.

Synthetic Aperture Radar (SAR) imagery is inherently marred by speckle noise, which undermines image quality and complicates subsequent analytical endeavors. While numerous strategies have been suggested in existing literature to mitigate this unwanted noise, the challenge of eliminating speckle while conserving subtle structural and textural details inherent in the raw data remains unresolved. In this article, we propose a comprehensive approach combining multi-domain analysis with gradient information processing for SAR. Our method aims to effectively suppress speckle noise while retaining crucial image characteristics. By leveraging multi-domain analysis techniques, we exploit both spatial and frequency domain information to gain a deeper insight into image structures. Additionally, we introduce a novel gradient information processing step that utilizes local gradient attributes to guide the process. Experimental results obtained from synthetic and real SAR imagery illustrate the effectiveness of our approach in terms of speckle noise reduction and preservation of image features. Quantitative assessments demonstrate substantial enhancements in image quality, indicating superior performance compared to current state-of-the-art methods.

Image restoration is a fundamental low-level vision task in consumer electronics and edge computing systems, aiming to recover high-quality images from degraded observations. While convolutional neural networks (CNNs) based restoration algorithms excel at local feature modeling, their limited receptive fields hinder global dependency learning. In contrast, Transformer-based methods enable global contextual reasoning but often incur prohibitive complexity due to dense self-attention. To overcome these limitations, we propose HSR-Former, a Hybrid Sparse Range Attention Transformer for unified multi-degradation image restoration. HSR-Former incorporates a Spatial Feature Attention Unit (SFAU) to preserve fine-grained structural fidelity and introduces a Sparse Region-Focused Multi-Head Attention (SRF-MHA) for enabling adaptive long-range dependency modeling with sub-quadratic complexity. Moreover, a Multi-Branch Depthwise Feed-Forward Network (MBD-FFN) is introduced to enable multi-scale token interaction and accelerate optimization. Extensive experiments on standard benchmarks for denoising, low-light enhancement, and deblurring demonstrate that HSR-Former achieves a favorable trade-off between restoration accuracy and computational complexity. The source code of HSR-Former is publicly available at: https://github.com/debashis15/HSR-Former.

Image noise, often resulting from disturbances during image acquisition or imperfections in the imaging device, notably degrades the quality of digital images. The challenge of removing this noise has been addressed through various techniques, from traditional filtering and prior-based methods to more recent deep learning approaches. In this paper, we introduce DRSFANet, an advanced dual-path convolutional neural network (CNN) specifically designed to tackle both synthetic Additive White Gaussian Noise (AWGN) and real-world noise. DRSFANet incorporates several innovative components: a residual feature extraction module (FEB) equipped with dilated convolutional layers to enhance the receptive field and mitigate gradient vanishing issues, and novel attention modules—Frequency-Plane Attention Block (FPAB) and Residual Attention Block (RAB)—which improve feature extraction in both frequency and spatial domains. Furthermore, the model features a downsampling (DS) block that effectively consolidates essential features prior to their integration into subsequent network stages. Comprehensive experimental evaluations reveal that DRSFANet outperforms several state-of-the-art denoising methods, demonstrating superior performance in both synthetic and real datasets through rigorous quantitative and qualitative analysis.

As a result of the increasing need for capturing and processing visual data of the Earth’s surface, Synthetic Aperture Radar (SAR) technology has been widely embraced by all space research organisations. The primary drawback in the acquired SAR visuals (images) is the presence of unwanted granular noise, called “speckle”, which poses a limitation to their processing and analysis. Therefore removing this unwanted speckle noise from the captured SAR visuals, a process known as despeckling, becomes an important task. This article introduces a new despeckling residual network named SAR-CDCFRN. This network simultaneously extracts speckle components from both the spatial and inverse spatial channels. The extracted features are then correlated by a dual-layer attention block and further processed to predict the distribution of speckle in the input noisy image. The predicted distribution, which is the residual noise, is then mapped with the input noisy SAR data to generate a despeckled output image. Experimental results confirm the superiority of the proposed despeckling model over other existing technologies in the literature.

Image noise, commonly introduced during the acquisition process, significantly degrades visual quality and adversely affects downstream image processing tasks. To address this challenge while preserving fine structural details, we propose GIADNet: a Gradient-Inspired Attention-Driven Denoising Network. The proposed framework integrates gradient-guided feature enhancement, multi-scale representation learning, and attention-based refinement to achieve a superior balance between noise suppression and detail retention. In particular, the gradient information of the noisy input is fused with deep features early in the pipeline to enrich semantic representation. Furthermore, we introduce two dedicated modules: the Multi-Pooling Pixel Attention (MPPA) module, which adaptively emphasizes informative pixels, and the Multi-Scale Attention Block (MSAB), designed to capture hierarchical contextual dependencies across varying spatial resolutions. Extensive experiments on standard benchmarks demonstrate that GIADNet achieves highly competitive performance, surpassing several state-of-the-art methods in both quantitative metrics and visual quality. Ablation studies further validate the effectiveness of each component, underscoring the importance of our attention-guided multi-scale design in advancing the field of image denoising. Code is available at: https://github.com/debashis15/GIADNet.

Blur and noise are intrinsic factors that develop during the image acquisition process and poses a limitation on the image acquisition system. Denoising and deconvolution are a class of image processing algorithms that aim to recover these spatially blurred and noisy images. Both the algorithms have their own merits and drawbacks, with denoising primarily focusing on noise removal and deconvolution on blur removal. In this paper, we propose a variational formulation for applying denoising and deconvolution jointly on an acquired image. We make use of the nuclear norm of Hessian as the regularizer and solve the resultant optimization problem using primal dual hybrid gradient descent algorithms. Comparison with multiple classical techniques, for different image degradation scenarios, justifies the merit of using such a joint denoising-deconvolution scheme for image restoration.

The interpretation of remote sensing images remains a significant challenge due to their complex, information-rich nature. Current remote sensing visual question answering (RSVQA) techniques have been a step forward toward building intelligent analysis systems for remote sensing images. However, most existing RSVQA models that rely on ResNet, VGG, and Swin transformers as visual feature extractors often fail to capture complex visual relationships, particularly the intricate dependencies between segmented regions and depth-related features in remote sensing data. To address these limitations, this letter introduces a novel RSVQA approach that leverages state-of-the-art components with an innovative architecture to advance interactive remote sensing analysis. The proposed model features a novel dual-layer visual attention mechanism in the representation module to process intricate features and capture regional relationships alongside processing the overall features. The fusion module employs a unique attention-based design, combining both self-attention and mutual attention, to integrate these features into a unified vector representation. Finally, the answering module utilizes a refined multilayer perceptron classifier for precise response generation. Evaluations on an RSVQA benchmark demonstrate the system's superiority over existing methods, marking a significant step forward in remote sensing analytics.

Hyperspectral imaging (HSI) involves capturing a scene across a broad spectrum of electromagnetic wavelengths to facilitate detailed analysis and predictive applications. Nev-ertheless, noise introduced during the acquisition process poses a significant challenge for the imaging system. In hyperspectral imaging (HSI), this noise generally consists of a blend of Gaussian and impulse types, making efficient noise reduction, or denoising, crucial for accurate image analysis. This study proposes a maximum a posteriori (MAP) based denoising solution, for eliminating the presence of mixed Gaussian-impulse noise in HSI, by constructing a joint optimization model for both noise distributions. The proposed solution utilizes the primal-dual hybrid gradient descent (PDHG) algorithm for solving and minimizing the joint optimization model. Rigorous tests on both synthetic and real hyperspectral imaging datasets demonstrate the efficacy of the suggested noise reduction technique.

Synthetic Aperture Radar (SAR) technology stands at the forefront of capturing and processing Earth's surface visuals due to its widespread acceptance across various organizations. However, the presence of unwanted random granular interference, commonly referred to as "speckle," poses a significant challenge in SAR data processing. Addressing this challenge, known as "despeckling," is crucial for extracting clear SAR visuals. This article introduces a novel CNN-based approach for despeckling SAR visuals contaminated with speckle. Our proposed model integrates a Generative Adversarial Network (GAN) module to estimate the distribution of contaminating speckle components from the input SAR data. Concurrently, a gradient estimator module captures the crisp changes in textural information within the input data. Subsequently, the input SAR data, the estimated noise distribution, and the extracted gradient undergo further processing through a deep convolutional module to generate a clean SAR visual. Unlike traditional methods that focus solely on learning the residual noisy component or the clean data, our proposed despeckling model learns the degradation pattern caused by noisy components while emphasizing gradient information, thereby capturing critical minute information. Experimental results demonstrate that our methodology significantly enhances despeckling performance compared to existing technologies in the literature. This research presents a promising step forward in advancing SAR visual despeckling techniques, with implications for improved data quality and interpretation in various applications.

Low-light images have less contrast and blurry details, which make them difficult for humans to see and difficult for computer vision algorithms to process. Traditional restoration techniques often fall short in retaining fine details and delivering visually appealing outcomes due to their reliance on global adjustments and manually designed features. While recent advancements in deep learning have shown promise in this area, they still struggle to effectively manage the wide range of illumination conditions encountered in low-light scenarios. To address these challenges, we propose “EnlightenNet, ‘’ a novel multi-channel deep learning framework specifically designed for low-light image restoration. The framework utilizes a dual-channel strategy: the first channel directly extracts intrinsic features from the input image, whereas the second channel captures features from a negatively logarithmic transformed feature space. Both channels operate within a unified Convolutional Neural Network (CNN) structure, which includes a shallow Local Feature Encoder (LFE) followed by a purposefully designed Multi-Channel Feature Dependency Encoder (MCFDE). The outputs from both channels are combined, and another Local Feature Encoder (LFE) is applied to reconstruct the enhanced image. By integrating logarithmic transformation in one channel and direct feature extraction in the other, the model can capture subtle details under low-light conditions. Significant experimental comparisons with cutting-edge techniques demonstrate the efficacy of the proposed model.

In the realm of computer vision, image denoising remains a formidable challenge with profound implications for fields like medical imaging, remote sensing, and photography. Despite notable advancements in deep learning, there are enduring challenges: current convolutional neural networks (CNNs) frequently struggle with training complexities due to their emphasis on increased network depth. At the same time, these networks often fail to adequately consider the crucial role of gradient information in the denoising process. Furthermore, there is a distinct gap in leveraging transform domain analysis in image denoising. This study addresses these limitations with MDFIDNet, a novel triple-phase attentive fusion network tailored for image denoising. MDFIDNet integrates three independent feature extraction pipelines: a frequency domain processing pipeline (FDP) enhanced by a multi-scale convolutional attention Block (MSCAB), a spatial domain processing pipeline (SDP) focusing on detail feature preservation, and a gradient-domain processing pipeline (GDP) driven by multidirectional gradient information. Experimental validation demonstrates that MDFIDNet surpasses existing benchmarks, exhibiting robust performance across diverse datasets. Comprehensive ablation studies underscore the individual contributions of each network component, elucidating the novel advancements that underpin MDFIDNet's superior denoising efficacy. The source code and further details are available in the https://github.com/debashis15/MDFIDNet.

The internet is the prime source to watch movies and micro-videos on platforms like YouTube, Netflix and many popular websites. All these online platforms are query-based search engines which extends a burden to the user to search and find a movie or video of their choice. The problem can be solved by developing better video recommender systems that will assist users in finding more helpful content and improving their overall experience. Deep learning is the leading solution for a large volume of multimedia data for personalized recommendations based on user interests. Feature-based solutions for video recommendation systems can be broadly classified under seven different categories: 1) User Embeddings – determining a user's specific interests, 2) Representation of Item – the user's dynamic interest based on their historically accessed items, 3) neighbour-assisted representation – we find similar users history data for generating Neighbour (history) interest information, 4) Categorical representation – It is learned by classifying the user's historical items into distinct categories and recognizing their differences, 5) Collaborative representation, 6) Hybrid representation – While neighbour-assisted characterization, which defines user profile from a collaborative perspective, characterises user interest from a customised perspective at the item and category level, and 7) Using rich contents (e.g., scene, meta, motion etc.) – uses to overcome restrictions caused by the absence of specific ones. In this book chapter, we will focus in detail on the principles and deep learning solutions that exist for online video recommendations to users. We will cover in detail the overview and the literature and shall also include an experimental analysis section wherein we shall analyze the performance of the various video recommender systems on different multimedia datasets.

In this paper, we propose a variational approach towards denoising magnetic resonance images (MRI) corrupted by spatially variant and signal-dependent Rician noise in a deep learning framework. To obtain a mathematically sound inference network, approximate variational posteriors are designed keeping in mind the Rician nature of noise. The proposed work tackles the denoising problem in several different ways. Firstly, the prior assumption on data in the variational posterior is motivated by the heavy-tailed marginal distribution of image gradients in natural images. This is captured by the sparsity promoting hyper-Laplacian prior on MR data. Similarly, median absolute deviation under Gaussian prior helps in the estimation of noise in the variational lower bound of marginal log likelihood term. Secondly, noise estimation from the background regions of the noisy data under the assumption of Rayleigh distribution prevents the addition of extra sub-network for the estimation of spatially variant noise level parameters. Thirdly, feature-wise transformation of intermediate layers is performed using anatomical planes segmentation maps (APSM) for context-based network conditioning. Here, affine transformation parameters generated from APSM are modulated with the input features for spatial feature transformation. Fourthly, to capture the long-range dependencies lost in deeper convolutional layers, multiscale global feature fusion block (GFFuB) is used. Lastly, experimental results over synthetically corrupted MR data and real data obtained from MR scanners suggest the potential utility of the proposed model in real time.

Mixed noise frequently affects optical images, complicating their interpretation and analysis. Among the most prevalent types of mixed noise are Gaussian and impulse noise. which can significantly degrade image quality. This study proposes a novel Bayesian method for removing mixed Gaussian-Impulse noise from optical images. The approach utilizes total variation (TV) and the nuclear norm of the Hessian matrix as regularization parameters within an optimization framework. These parameters are derived from maximum a posteriori (MAP) estimations of the noise statistics. TV regularization ensures the smoothness of the solution while incorporating the Hessian matrix, which helps preserve fine details in the final optimized image. The problem is then addressed using primal-dual algorithms, which efficiently solve the proposed optimization problem. Experimental results demonstrate that the proposed method significantly improves image restoration quality compared to existing denoising techniques. The study's findings indicate that with its sophisticated regularization and optimization strategies. this Bayesian approach offers a robust solution for effectively denoising optical images contaminated with mixed Gaussian-Impulse noise.

Continual occurrences of frequency interference result in the emergence of unpredictable speckle patterns within the unprocessed data gathered by Synthetic Aperture Radar (SAR) sensors. The overall quality of the acquired images are noticeably diminished because of this unwanted speckle which overlays a granular texture onto the visuals. Many methods have been suggested in scientific literature to eliminate this undesirable speckle. Nevertheless, the challenge of eliminating speckles while retaining subtle structural and textural details in the original data remains a challenging task to this date. In this paper, we have introduced a distinctive auto-encoder driven SAR despeckling model with three main modules: Despeckling Encoder (DE) for despeckling, Speckle Extractor (SE) for estimating the speckle distribution, and Reconstruction Module (RM) for reconstructing the noisy image. The DE processes both spatial and frequency components, while the SE and RM handle the speckle noise and its impact on the clear image, thereby fine-tuning the DE module. The comprehensive experimental simulations provided in this study offer robust evidence of the superior performance of the proposed model when compared to numerous classical and contemporary methods outlined in the existing literature.

Images obtained from hyperspectral sensors provide information about the target area that extends beyond the visible portions of the electromagnetic spectrum. However, due to sensor limitations and imperfections during the image acquisition and transmission phases, noise is introduced into the acquired image, which can have a negative impact on downstream analyses such as classification, target tracking, and spectral unmixing. Noise in hyperspectral images (HSI) is modelled as a combination from several sources, including Gaussian/impulse noise, stripes, and deadlines. An HSI restoration method for such a mixed noise model is proposed. First, a joint optimisation framework is proposed for recovering hyperspectral data corrupted by mixed Gaussian‐impulse noise by estimating both the clean data as well as the sparse/impulse noise levels. Second, a hyper‐Laplacian prior is used along both the spatial and spectral dimensions to express sparsity in clean image gradients. Third, to model the sparse nature of impulse noise, an ℓ1 − norm over the impulse noise gradient is used. Because the proposed methodology employs two distinct priors, the authors refer to it as the hyperspectral dual prior (HySpDualP) denoiser. To the best of authors' knowledge, this joint optimisation framework is the first attempt in this direction. To handle the non‐smooth and non‐convex nature of the general ℓp − norm‐based regularisation term, a generalised shrinkage/thresholding (GST) solver is employed. Finally, an efficient split‐Bregman approach is used to solve the resulting optimisation problem. Experimental results on synthetic data and real HSI datacube obtained from hyperspectral sensors demonstrate that the authors’ proposed model outperforms state‐of‐the‐art methods, both visually and in terms of various image quality assessment metrics.