Global Aligned Structured Sparsity Learning for Efficient Image Super-Resolution.

Multi-scale information distillation network for efficient image super-resolution

Deploying advanced Machine Learning (ML), particularly Deep Neural Networks (DNNs), on resource-constrained edge devices is crucial for realizing low-latency, privacy-preserving, and reliable edge intelligence applications. However, a significant gap exists between the high computational, memory, and energy demands of state-of-the-art models and the severe limitations inherent to edge hardware. This review systematically analyzes the field of lightweight ML for edge devices, aiming to bridge this gap. Methods: Focusing on the inference phase, the review critically examines three primary technical pillars: (1) Model Compression techniques, including knowledge distillation, network pruning (structured and unstructured), and quantization; (2) Efficient Neural Architecture Design of inherently compact models (e.g., MobileNet, ShuffleNet, EfficientNet series); and (3) Hardware-aware Optimization and Adaptation, encompassing operator fusion, dedicated inference engines, and leveraging heterogeneous systems. ​Results and Conclusion: The analysis highlights key achievements in reducing model size, complexity, and latency while maintaining accuracy. However, fundamental challenges persist, including the accuracy-efficiency tradeoff, hardware fragmentation, the memory wall bottleneck, and privacy/security concerns during deployment. Emerging solutions like neural-symbolic learning, adaptive federated learning, hardware-aware Neural Architecture Search (NAS), Processing-in-Memory (PIM) accelerators, and cross-stack co-design frameworks represent promising future directions. Overcoming these challenges is strategically vital for unlocking the full potential of ubiquitous, real-time edge intelligence.

Deep Neural Networks are often vulnerable to adversarial attacks. Neural Architecture Search (NAS), one of the tools for developing novel deep neural architectures, demonstrates superior performance in prediction accuracy in various machine learning applications. However, the performance of a neural architecture discovered by NAS against adversarial attacks has not been sufficiently studied, especially under the regime of knowledge distillation. Given the presence of a robust teacher, we investigate if NAS would produce a robust neural architecture by inheriting robustness from the teacher. In this paper, we propose Robust Neural Architecture Search by Cross-Layer knowledge distillation (RNAS-CL), a novel NAS algorithm that improves the robustness of NAS by learning from a robust teacher through cross-layer knowledge distillation. Unlike previous knowledge distillation methods that encourage close student-teacher output only in the last layer, RNAS-CL automatically searches for the best teacher layer to supervise each student layer. Experimental results demonstrate the effectiveness of RNAS-CL and show that RNAS-CL produces compact and adversarially robust neural architectures. Our results point to new approaches for finding compact and robust neural architecture for many applications. The code of RNAS-CL is available at https://github.com/Statistical-Deep-Learning/RNAS-CL.

With the rapid growth of Internet video image information, there is a large amount of redundancy in image data. Use less data stream information to transfer the image or the amount of information contained in the image. Its purpose is to reduce the redundancy of images, so as to store them at low bit rate and reduce the data storage space. In the general image compression method, the hybrid coding framework is adopted. Each algorithm adopts a fixed algorithm mode through a specific design algorithm, without global optimization. Image compression is mainly divided into prediction, transformation, quantization, digital entropy coding and other steps. At present, there are many researches on super-resolution network based on deep learning technology. The main function is to reconstruct high-resolution image replace image magnification low-resolution images such as linear interpolation, which has a great performance improvement image resolution, noise reduction, deblurring and so on, but there is no effective way to use super-resolution network applications to improve quality of compression reconstructed image quality. This paper involves a new method that using image super-resolution residual learning network to improve quality of compression image, our method, the reduced image is encoded into a content stream and a transmission corresponding parameter is encoded into a model stream. Firstly, the original image is scaled down 1/2 size of source image, then encode the small image into content stream with the existing codec. Secondly, the residual learning super-resolution (SR) network is used for image filtering to scale up reconstructed image with decode image resizing method and boost the quality of edge feature extraction of image. Our results show that there is significant performance improvement of h265 in low resolution reconstructed image (bits-per-pixel less than 0.1).

Neural architecture search (NAS) targets at finding the optimal architecture of a neural network for a problem or a family of problems. Evaluations of neural architectures are very time-consuming. One of the possible ways to mitigate this issue is to use low-fidelity evaluations, namely training on a part of a dataset, fewer epochs, with fewer channels, etc. In this paper, we propose a Bayesian multi-fidelity method for neural architecture search: MF-KD. The method relies on a new approach to low-fidelity evaluations of neural architectures by training for a few epochs using a knowledge distillation. Knowledge distillation adds to a loss function a term forcing a network to mimic some teacher network. We carry out experiments on CIFAR-10, CIFAR-100, and ImageNet-16-120. We show that training for a few epochs with such a modified loss function leads to a better selection of neural architectures than training for a few epochs with a logistic loss. The proposed method outperforms several state-of-the-art baselines.

Recent AI research has witnessed increasing interests in automatically designing the architecture of deep neural networks, which is coined as neural architecture search (NAS). The automatically searched network architectures via NAS methods have outperformed manually designed architectures on some NLP tasks. However, training a large number of model configurations for efficient NAS is computationally expensive, creating a substantial barrier for applying NAS methods in real-life applications. In this paper, we propose to accelerate neural architecture search for natural language processing based on knowledge distillation (called KD-NAS). Specifically, instead of searching the optimal network architecture on the validation set conditioned on the optimal network weights on the training set, we learn the optimal network by minimizing the knowledge loss transferred from a pre-trained teacher network to the searching network based on Earth Mover's Distance (EMD). Experiments on five datasets show that our method achieves promising performance compared to strong competitors on both accuracy and searching speed. For reproducibility, we submit the code at: https://github.com/lxk00/KD-NAS-EMD.

Deep learning techniques have recently demonstrated remarkable success in numerous domains. Typically, the success of these deep learning models is measured in terms of performance metrics such as accuracy and mean average precision (mAP). Generally, a model's high performance is highly valued, but it frequently comes at the expense of substantial energy costs and carbon footprint emissions during the model building step. Massive emission of CO2 has a deleterious impact on life on earth in general and is a serious ethical concern that is largely ignored in deep learning research. In this article, we mainly focus on environmental costs and the means of mitigating carbon footprints in deep learning models, with a particular focus on models created using knowledge distillation (KD). Deep learning models typically contain a large number of parameters, resulting in a 'heavy' model. A heavy model scores high on performance metrics but is incompatible with mobile and edge computing devices. Model compression techniques such as knowledge distillation enable the creation of lightweight, deployable models for these low-resource devices. KD generates lighter models and typically performs with slightly less accuracy than the heavier teacher model (model accuracy by the teacher model on CIFAR 10, CIFAR 100, and TinyImageNet is 95.04%, 76.03%, and 63.39%; model accuracy by KD is 91.78%, 69.7%, and 60.49%). Although the distillation process makes models deployable on low-resource devices, they were found to consume an exorbitant amount of energy and have a substantial carbon footprint (15.8, 17.9, and 13.5 times more carbon compared to the corresponding teacher model). The enormous environmental cost is primarily attributable to the tuning of the hyperparameter, Temperature (τ). In this article, we propose measuring the environmental costs of deep learning work (in terms of GFLOPS in millions, energy consumption in kWh, and CO2 equivalent in grams). In order to create lightweight models with low environmental costs, we propose a straightforward yet effective method for selecting a hyperparameter (τ) using a stochastic approach for each training batch fed into the models. We applied knowledge distillation (including its data-free variant) to problems involving image classification and object detection. To evaluate the robustness of our method, we ran experiments on various datasets (CIFAR 10, CIFAR 100, Tiny ImageNet, and PASCAL VOC) and models (ResNet18, MobileNetV2, Wrn-40-2). Our novel approach reduces the environmental costs by a large margin by eliminating the requirement of expensive hyperparameter tuning without sacrificing performance. Empirical results on the CIFAR 10 dataset show that the stochastic technique achieves an accuracy of 91.67%, whereas tuning achieves an accuracy of 91.78%-however, the stochastic approach reduces the energy consumption and CO2 equivalent each by a factor of 19. Similar results have been obtained with CIFAR 100 and TinyImageNet dataset. This pattern is also observed in object detection classification on the PASCAL VOC dataset, where the tuning technique performs similarly to the stochastic technique, with a difference of 0.03% mAP favoring the stochastic technique while reducing the energy consumptions and CO2 emission each by a factor of 18.5.

As convolutional neural networks (CNNs) become gradually larger and deeper, their applicability in real-time and resource-constrained environments is significantly limited. Furthermore, while self-attention (SA) mechanisms excel at capturing global dependencies, they often emphasize low-frequency information and struggle to represent fine local details. To overcome these limitations, we propose a multi-scale adaptive modulation network (MAMN) for image super-resolution. The MAMN mainly consists of a series of multi-scale adaptive modulation blocks (MAMBs), each of which incorporates a multi-scale adaptive modulation layer (MAML), a local detail extraction layer (LDEL), and two Swin Transformer Layers (STLs). The MAML is designed to capture multi-scale non-local representations, while the LDEL complements this by extracting high-frequency local features. Additionally, the STLs enhance long-range dependency modeling, effectively expanding the receptive field and integrating global contextual information. Extensive experiments demonstrate that the proposed method achieves an optimal trade-off between computational efficiency and reconstruction performance across five benchmark datasets.

Adversarially trained deep neural networks have significantly improved performance of single image super resolution, by hallucinating photorealistic local textures, thereby greatly reducing the perception difference between a real high resolution image and its super resolved (SR) counterpart. However, application to medical imaging requires preservation of diagnostically relevant features while refraining from introducing any diagnostically confusing artifacts. We propose using a deep convolutional super resolution network (SRNet) trained for (i) minimising reconstruction loss between the real and SR images, and (ii) maximally confusing learned relativistic visual Turing test (rVTT) networks to discriminate between (a) pair of real and SR images (T1) and (b) pair of patches in real and SR selected from region of interest (T2). The adversarial loss of T1 and T2 while backpropagated through SRNet helps it learn to reconstruct pathorealism in the regions of interest such as white blood cells (WBC) in peripheral blood smears or epithelial cells in histopathology of cancerous biopsy tissues, which are experimentally demonstrated here. Experiments performed for measuring signal distortion loss using peak signal to noise ratio (pSNR) and structural similarity (SSIM) with variation of SR scale factors, impact of rVTT adversarial losses, and impact on reporting using SR on a commercially available artificial intelligence (AI) digital pathology system substantiate our claims.

Efficient deep learning based methods have achieved significant performance in single image super-resolution. Recent research on efficient super-resolution has mainly focused on reducing the number of parameters and computational complexity through various network designs, enabling models to be better deployed on resource constrained devices. In this work, we propose a novel and effective super-resolution model based on attention mechanism and structural re-parameterization, called Dual Path Attention and Re-parameterization Network DPARN , which uses a hybrid attention mechanisms to balance the running speed and reconstruction quality of the model. Specifically, we utilize grouped convolution to introduce both parameter free and enhanced spatial attention to improve the feature extraction capability of student networks. Meanwhile, we ado-pted a novel lightweight network training strategy that first uses knowledge distillation for initial training, during which structured knowledge from the teacher network is transmitted to the student network. Then, multiple loss functions are combined to fine-tuning the student network, in order to preserve high-frequency details and avoid excessive smoothing caused by pixel loss. Finally, extensive experiments conducted on four benchmark datasets demonstrated the effectiveness and efficiency of proposed DPARN. Our method achieves PSNR SSIM performance comparable to state-of-the-art efficient super-resolution models, with faster inference speed and fewer network parameters.

Efficient image super-resolution with semantic guidance and denoising modules

Abstract. As deep learning models advance, their use in Super Resolution (SR) tasks has become pivotal for enhancing remote sensing low-resolution (LR) satellite images. However, the decision making processes within these models remain opaque, especially in remote sensing applications where transparency is critical. This paper focuses on applying Explainable Artificial Intelligence (XAI) techniques, particularly Local Attribution Maps (LAMs), to analyze and interpret the internal behavior of both general purpose and remote sensing specific SR neural networks. General purpose models like Generative Adversarial Network for Super Resolution (SRGAN), Enhanced Deep Super-Resolution Network (EDSR), Efficient Super-Resolution Transformer (ESRT), and Hybrid-Attention Transformer (HAT), although highly effective in SR tasks, were originally designed for broader image enhancement challenges; in contrast, Hybrid-Scale Self-Similarity Exploitation Network (HSENet) and Multi-scale Enhanced Network (MEN) are tailored for the unique complexities of remote sensing, such as varied textures and intricate scene features. By leveraging LAMs, we highlight how different networks prioritize and process features such as edges, textures, and high frequency details to generate super resolved outputs. The comparative study between general purpose and remote sensing specific networks outlines each model’s strengths and weaknesses in managing the data present in remote sensing imagery. This approach addresses the previously unexplored application of LAMs in remote sensing for SR, contributing to the research in this field and providing deeper insights into the interpretability and transparency of widely used SR models. Furthermore, by drawing parallels between feature attribution in SR and classification tasks, we suggest new pathways for integrating semantic information to refine model transparency and performance in remote sensing applications.

Nonhomogeneous haze removal is a challenging problem, which does not follow the physical scattering model of haze. Numerous existing methods focus on homogeneous haze removal by generating transmission map of the image, which is not suitable for nonhomogeneous dehazing tasks. Some methods use end-to-end model but are also designed for homogeneous haze. Inspired by Knowledge Transfer Dehazing Network and Trident Dehazing Network, we propose a model with super-resolution method and knowledge transfer method. Our model consists of a teacher network, a dehaze network and a super-resolution network. The teacher network provides the dehaze network with reliable prior, the dehaze network focuses primarily on haze removal, and the super-resolution network is used to capture details in the hazy image. Ablation study shows that the super-resolution network has significant benefit to image quality. And comparison shows that our model out- performs previous state-of-the-art methods in terms of perceptual quality on NTIRE2021 NonHomogeneous Dehazing Challenge dataset, and also performs well on other datasets.

Knowledge distillation has become increasingly important in model compression. It boosts the performance of a miniaturized student network with the supervision of the output distribution and feature maps from a sophisticated teacher network. Some recent works introduce multi-teacher distillation to provide more supervision to the student network. However, the effectiveness of multi-teacher distillation methods are accompanied by costly computation resources. To tackle with both the efficiency and the effectiveness of knowledge distillation, we introduce the feature aggregation to imitate the multi-teacher distillation in the single-teacher distillation framework by extracting informative supervision from multiple teacher feature maps. Specifically, we introduce DFA, a two-stage Differentiable Feature Aggregation search method that motivated by DARTS in neural architecture search, to efficiently find the aggregations. In the first stage, DFA formulates the searching problem as a bi-level optimization and leverages a novel bridge loss, which consists of a student-to-teacher path and a teacher-to-student path, to find appropriate feature aggregations. The two paths act as two players against each other, trying to optimize the unified architecture parameters to the opposite directions while guaranteeing both expressivity and learnability of the feature aggregation simultaneously. In the second stage, DFA performs knowledge distillation with the derived feature aggregation. Experimental results show that DFA outperforms existing distillation methods on CIFAR-100 and CINIC-10 datasets under various teacher-student settings, verifying the effectiveness and robustness of the design.KeywordsKnowledge distillationFeature aggregationDifferentiable architecture search

In this work, a simple and compact-structured convolutional neural network (CNN) entitled as Denoising and Super-Resolution (DnSR) network for single image denoising and super resolution (SISR) is presented. Using sparsity property (a smaller number of large intensity values and a larger number of small intensity values), the training of DnSR is made to converge early. We obtain the sparsity by learning residuals of image pixels instead of learning raw image pixels particularly for SR. We use a smaller receptive field of the image patch for faster training. We use gradient clipping to train DnSR net with an extremely large learning rate to overcome gradient inconsistency during training. An end-to-end mapping is done by learning the complex mapping between low-resolution (LR) image patches and the corresponding residual of high resolution (∆HR) image patches using DnSR net. The proposed DnSR net uses a smaller number of parameters (approximately 3 times lesser) than the existing state-of-the-art (SOTA) SRCNN method. The proposed DnSR provides superior reconstruction accuracy with less artifacts when compared to other SOTA methods. The performance of the DnSR is assessed qualitatively and quantitatively using different image quality metrics.