Lightweight Deep Neural Network Research Articles

Toward DNN of LUTs: Learning Efficient Image Restoration With Multiple Look-Up Tables.

The widespread usage of high-definition screens on edge devices stimulates a strong demand for efficient image restoration algorithms. The way of caching deep learning models in a look-up table (LUT) is recently introduced to respond to this demand. However, the size of a single LUT grows exponentially with the increase of its indexing capacity, which restricts its receptive field and thus the performance. To overcome this intrinsic limitation of the single-LUT solution, we propose a universal method to construct multiple LUTs like a neural network, termed MuLUT. First, we devise novel complementary indexing patterns, as well as a general implementation for arbitrary patterns, to construct multiple LUTs in parallel. Second, we propose a re-indexing mechanism to enable hierarchical indexing between cascaded LUTs. Finally, we introduce channel indexing to allow cross-channel interaction, enabling LUTs to process color channels jointly. In these principled ways, the total size of MuLUT is linear to its indexing capacity, yielding a practical solution to obtain superior performance with the enlarged receptive field. We examine the advantage of MuLUT on various image restoration tasks, including super-resolution, demosaicing, denoising, and deblocking. MuLUT achieves a significant improvement over the single-LUT solution, e.g., up to 1.1 dB PSNR for super-resolution and up to 2.8 dB PSNR for grayscale denoising, while preserving its efficiency, which is 100× less in energy cost compared with lightweight deep neural networks.

SteriCNN: Cloud native stego content sterilization framework

Modern robust steganography-based cyber attacks often bypass intrinsic cloud security measures, and contemporary steganalysis methods struggle to address these covert threats due to recent advancements in deep learning (DL)-based steganography techniques. Existing steganography removal methods are constrained by trade-offs involving high processing times, poor quality of sanitized images, and insufficient removal of steganographic content. This paper introduces SteriCNN, a lightweight deep residual neural network model designed for steganography removal. SteriCNN effectively eliminates embedded steganographic information while preserving the visual integrity of the sanitized images. We employ a series of convolutional blocks with three residual connections for feature extraction, feature learning, feature attention, and image reconstruction from the residue. The proposed model utilizes the correlation of channel features to achieve a faster learning rate, and by varying the dilation rate in convolutional blocks, the model achieves wider receptive fields, enabling it to cover larger areas of the input image at each layer. SteriCNN is targeted for blind image sterilization for real-time use cases due to its low training and prediction time costs. Our study shows impressive results for both traditional and deep learning-based stego vulnerabilities, with approximately 90% of steganograms eliminated while maintaining an average PSNR value of 46 dB and an SSIM of 0.99 when tested with popular steganography methods.

Lightweight Advanced Deep Neural Network (DNN) Model for Early-Stage Lung Cancer Detection.

Background: Lung cancer, also known as lung carcinoma, has a high mortality rate; however, an early prediction helps to reduce the risk. In the current literature, various approaches have been developed for the prediction of lung carcinoma (at an early stage), but these still have various issues, such as low accuracy, high noise, low contrast, poor recognition rates, and a high false-positive rate, etc. Thus, in this research effort, we have proposed an advanced algorithm and combined two different types of deep neural networks to make it easier to spot lung melanoma in the early phases. Methods: We have used WDSI (weakly supervised dense instance-level lung segmentation) for laborious pixel-level annotations. In addition, we suggested an SS-CL (deep continuous learning-based deep neural network) that can be applied to the labeled and unlabeled data to improve efficiency. This work intends to evaluate potential lightweight, low-memory deep neural net (DNN) designs for image processing. Results: Our experimental results show that, by combining WDSI and LSO segmentation, we can achieve super-sensitive, specific, and accurate early detection of lung cancer. For experiments, we used the lung nodule (LUNA16) dataset, which consists of the patients' 3D CT scan images. We confirmed that our proposed model is lightweight because it uses less memory. We have compared them with state-of-the-art models named PSNR and SSIM. The efficiency is 32.8% and 0.97, respectively. The proposed lightweight deep neural network (DNN) model archives a high accuracy of 98.2% and also removes noise more effectively. Conclusions: Our proposed approach has a lot of potential to help medical image analysis to help improve the accuracy of test results, and it may also prove helpful in saving patients' lives.

Employing Different Algorithms of Lightweight Convolutional Neural Network Models in Image Distortion Classification

The majority of applications use automatic image recognition technologies to carry out a range of tasks. Therefore, it is crucial to identify and classify image distortions to improve image quality. Despite efforts in this area, there are still many challenges in accurately and reliably classifying distorted images. In this paper, we offer a comprehensive analysis of models of both non-lightweight and lightweight deep convolutional neural networks (CNNs) for the classification of distorted images. Subsequently, an effective method is proposed to enhance the overall performance of distortion image classification. This method involves selecting features from the pretrained models’ capabilities and using a strong classifier. The experiments utilized the kadid10k dataset to assess the effectiveness of the results. The K-nearest neighbor (KNN) classifier showed better performance than the naïve classifier in terms of accuracy, precision, error rate, recall and F1 score. Additionally, SqueezeNet outperformed other deep CNN models, both lightweight and non-lightweight, across every evaluation metric. The experimental results demonstrate that combining SqueezeNet with KNN can effectively and accurately classify distorted images into the correct categories. The proposed SqueezeNet-KNN method achieved an accuracy rate of 89%. As detailed in the results section, the proposed method outperforms state-of-the-art methods in accuracy, precision, error, recall, and F1 score measures.

An approach for classification of breast cancer using lightweight deep convolution neural network

The rapid advancement of deep learning has generated considerable enthusiasm regarding its utilization in addressing medical imaging issues. Machine learning (ML) methods can help radiologists to diagnose breast cancer (BCs) barring invasive measures. Informative hand-crafted features are essential prerequisites for traditional machine learning classifiers to achieve accurate results, which are time-consuming to extract. In this paper, our deep learning algorithm is created to precisely identify breast cancers on screening mammograms, employing a training method that effectively utilizes training datasets with either full clinical annotation or solely the cancer status of the entire image. The proposed approach utilizes Lightweight Convolutional Neural Network (LWCNN) that allows automatic extraction features in an end-to-end manner. We have tested LWCNN model in two experiments. In the first experiment, the model was tested with two cases' original and enhancement datasets 1. It achieved 95 %, 93 %, 99 % and 98 % for training and testing accuracy respectively. In the second experiment, the model has been tested with two cases' original and enhancement datasets 2. It achieved 95 %, 91 %, 99 % and 92 % for training and testing accuracy respectively. Our proposed method, which uses various convolutional network to classify screening mammograms achieved exceptional performance when compared to other methods. The findings from these experiments clearly indicate that automatic deep learning techniques can be trained effectively to attain remarkable accuracy across a wide range of mammography datasets. This holds significant promise for improving clinical tools and reducing both false positive and false negative outcomes in screening mammography.

Robust filter pruning guided by deep frequency-features for edge intelligence

With the popularity of mobile applications, smart terminals require the deployment of lightweight deep neural networks (DNNs). However, those lightweight DNNs deployed on resource-constrained edge devices, have serious security vulnerabilities to adversarial examples. Pruning methods have been widely used to obtain lightweight DNNs, which mainly focus on improving classification accuracy, but ignore robustness. To this end, we present a Fourier space-based robust pruning algorithm (FSRP). We find that a robust network pays more attention to the low-frequency region of the adversarial examples’ feature maps, but ignores the high-frequency region. According to this finding, we design a filter robust indicator, a ratio of low-frequency components to high-frequency components of the feature map, to guide the pruning process. With this new robust pruning criterion, we adopt a strategy of local pruning that removes the filters with low robustness layer by layer in the model. Extensive experiments on CIFAR10/CIFAR100 and ImageNet show that the robust pruning accuracy is significantly improved under the FSRP robust pruning criterion. On CIFAR10, a pruning ratio of 90% on the VGG16 network still shows a 14.1% improvement in robust accuracy under auto attacks (AA).

Ultra-lightweight convolution-transformer network for early fire smoke detection

BackgroundForests are invaluable resources, and fire is a natural process that is considered an integral part of the forest ecosystem. Although fire offers several ecological benefits, its frequent occurrence in different parts of the world has raised concerns in the recent past. Covering millions of hectares of forest land, these fire incidents have resulted in the loss of human lives, wild habitats, civil infrastructure, and severe damage to the environment. Around 90% of wildland fires have been caused by humans intentionally or unintentionally. Early detection of fire close to human settlements and wildlife centuries can help mitigate fire hazards. Numerous artificial intelligence-based solutions have been proposed in the past decade that prioritize the detection of fire smoke, as it can be caught through remote sensing and provide an early sign of wildland fire. However, most of these methods are either computationally intensive or suffer from a high false alarm rate. In this paper, a lightweight deep neural network model is proposed for fire smoke detection in images captured by satellites or other remote sensing sources.ResultsWith only 0.6 million parameters and 0.4 billion floating point operations per second, the hybrid network of convolutional and vision transformer blocks efficiently detects smoke in normal and foggy environmental conditions. It outperforms seven state-of-the-art methods on four datasets, including a self-collected dataset from the “Moderate Resolution Imaging Spectroradiometer” satellite imagery. The model achieves an accuracy of more than 99% on three datasets and 93.90% on the fourth dataset. The t-distributed stochastic neighbor embedding of extracted features by the proposed model demonstrates its superior feature learning capabilities. It is remarkable that even a tiny occurrence of smoke covering just 2% of the satellite image area is efficiently detected by the model.ConclusionsWith low memory and computational demands, the proposed model works exceedingly well, making it suitable for deployment in resource constrained devices for forest surveillance and early fire smoke detection.

A lightweight deep learning architecture for automatic modulation classification of wireless internet of things

AbstractThe wireless Internet of Things (IoT) is widely used for data transmission in power systems. Wireless communication is an important part of the IoT. The existing modulation classification algorithms have low classification accuracy when facing strong electromagnetic interference, which causes decoding error link interruption and wastes wireless channel resources. Therefore, it is necessary to study signal modulation classification methods in a low signal‐to‐noise ratio (SNR) environment. In this paper, a lightweight Deep Neural Networks (DNNs) modulation classification method based on the Informer architecture classifier and two‐dimensional (2‐D) curves input of the spectral correlation function (SCF) is proposed, which uses in‐phase and quadrature (I/Q) signals to generate 2‐D cross‐section SCF curve first and then feeds the feature curve into the Informer network to classify the modulation method. This model can better learn the robustness characteristics in a long sequence. Through testing, the classification accuracy of the modulation signal is not much lower than that of the current good classification method when the SNR is 10 dB, and this method can still show higher accuracy when hardware resources are limited. It is a compact design of a modulation classification model and easy to deploy on low‐cost embedded platforms.

Lightweight deep neural network for radio frequency interference detection and segmentation in synthetic aperture radar

Radio frequency interference (RFI) poses challenges in the analysis of synthetic aperture radar (SAR) images. Existing RFI suppression systems rely on prior knowledge of the presence of RFI. This paper proposes a lightweight neural network-based algorithm for detecting and segmenting RFI (LDNet) in the time-frequency domain. The network accurately delineates RFI pixel regions in time-frequency spectrograms. To mitigate the impact on the operational speed of the entire RFI suppression system, lightweight modules and pruning operations are introduced. Compared to threshold-based RFI detection algorithms, deep learning-based segmentation networks, and AC-UNet specifically designed for RFI detection, LDNet achieves improvements in mean intersection over union (MIoU) by 24.56%, 13.29%, and 7.54%, respectively.Furthermore, LDNet reduces model size by 99.03% and inference latency by 24.53% compared to AC-UNet.

Evaluating Dimensionality Reduction Methods for the Detection of Industrial IoT Attacks in Edge Computing

Edge computing is essential for 6G mobile networks for improving reliability, reducing data rates and latency, and enhancing mobile connectivity. Edge computing is also meant to meet the increasing demands of the Internet of Things (IoT)/ Internet of Everything (IoE). In these approaches, IIoT systems necessitate precision, reliability, and scalability, while vulnerabilities in IIoT systems may lead to financial losses and safety hazards. To tackle this, Edge AI/ML-based IDSs provide adaptability and robustness for IIoT security challenges. These solutions improve security levels, threat detection rates, and response times. However, main issues such as limited resources and the accuracy of attack detection rate are challenged nowadays. In this paper, we present contributions in proposing an intrusion detection system (IDS) for edge devices deploying a lightweight deep neural network to detect IIoT attacks. We present performance analysis of the typical dimensionality reduction methods and balancing data features of the Edge-IIoT dataset to get higher performance metrics than other state-of-the-art studies.

Cataract and glaucoma detection based on Transfer Learning using MobileNet

A serious eye condition called cataracts can cause blindness. Early and accurate cataract detection is the most effective method for reducing risk and averting blindness. The optic nerve head is harmed by the neurodegenerative condition known as glaucoma. Machine learning and deep learning systems for glaucoma and cataract detection have recently received much attention in research. The automatic detection of these diseases also depends on deep learning transfer learning platforms like VeggNet, ResNet, and MobilNet. The authors proposed MobileNetV1 and MobileNetV2 based on an optimized architecture building lightweight deep neural networks using depth-wise separable convolutions. The experiments used publicly available data sets with both cataract & normal and glaucoma & normal images, and the results showed that the proposed model had the highest accuracy compared to the other models.

Monitoring Berlin Heart EXCOR by computer vision: Apreliminary implementation and evaluation.

EXCOR Pediatric is one of the most commonly used ventricular assist devices (VAD) for small children; it requires visual inspection of the diaphragm movement to assess its operating status. Although this visual inspection can only be performed by trained medical professionals, it can also be attempted by the recent advances in computer vision technology. Movement of the diaphragm in the operating EXCOR VAD was recorded as movies and annotated frame-by-frame in three classes according to the state of the diaphragm: "fill," "mid," and "empty." Three models, MobileNetV3, EfficientNetV2, and MobileViT, were trained using the frames, and their performance was evaluated based on the accuracy and area under the receiver operating characteristic curve (AROC). A total of 152 movies were available from two participants. Only the 10 mL pumps were used. Ninety-eight movies were used for annotation and frame extraction, and 7949 frames per class were included in the analysis. The macro-average accuracies of the three models were 0.88, 0.91, and 0.93, and the AROC were 0.99, 0.98, and 0.99 for MobileNetV3, EfficientNetV2, and MobileViT, respectively. Image recognition models based on lightweight deep neural networks could detect the diaphragm state of EXCOR VAD with sufficient accuracy, although there were limited variations in the dataset. This suggests the potential of computer vision for the automated monitoring of the EXCOR diaphragm position.

Deep learning with convolution neural network detecting mesiodens on panoramic radiographs: comparing four models.

The aim of this study was to develop an optimal, simple, and lightweight deep learning convolutional neural network (CNN) model to detect the presence of mesiodens on panoramic radiographs. A total of 628 panoramic radiographs with and without mesiodens were used as training, validation, and test data. The training, validation, and test dataset were consisted of 218, 51, and 40 images with mesiodens and 203, 55, and 61 without mesiodens, respectively. Unclear panoramic radiographs for which the diagnosis could not be accurately determined and other modalities were required for the final diagnosis were retrospectively identified and employed as the training dataset. Four CNN models provided within software supporting the creation of neural network models for deep learning were modified and developed. The diagnostic performance of the CNNs was evaluated according to accuracy, precision, recall and F1 scores, receiver operating characteristics (ROC) curves, and area under the ROC curve (AUC). In addition, we used SHapley Additive exPlanations (SHAP) to attempt to visualize the image features that were important in the classifications of the model that exhibited the best diagnostic performance. A binary_connect_mnist_LeNet model exhibited the best performance of the four deep learning models. Our results suggest that a simple lightweight model is able to detect mesiodens. It is worth referring to AI-based diagnosis before an additional radiological examination when diagnosis of mesiodens cannot be made on unclear images. However, further revaluation by the specialist would be also necessary for careful consideration because children are more radiosensitive than adults.

3D mobile regression vision transformer for collateral imaging in acute ischemic stroke.

The accurate and timely assessment of the collateral perfusion status is crucial in the diagnosis and treatment of patients with acute ischemic stroke. Previous works have shown that collateral imaging, derived from CT angiography, MR perfusion, and MR angiography, aids in evaluating the collateral status. However, such methods are time-consuming and/or sub-optimal due to the nature of manual processing and heuristics. Recently, deep learning approaches have shown to be promising for generating collateral imaging. These, however, suffer from the computational complexity and cost. In this study, we propose a mobile, lightweight deep regression neural network for collateral imaging in acute ischemic stroke, leveraging dynamic susceptibility contrast MR perfusion (DSC-MRP). Built based upon lightweight convolution and Transformer architectures, the proposed model manages the balance between the model complexity and performance. We evaluated the performance of the proposed model in generating the five-phase collateral maps, including arterial, capillary, early venous, late venous, and delayed phases, using DSC-MRP from 952 patients. In comparison with various deep learning models, the proposed method was superior to the competitors with similar complexity and was comparable to the competitors of high complexity. The results suggest that the proposed model is able to facilitate rapid and precise assessment of the collateral status of patients with acute ischemic stroke, leading to improved patient care and outcome.

Deep learning for 3D cephalometric landmarking with heterogeneous multi-center CBCT dataset.

Cephalometric analysis is critically important and common procedure prior to orthodontic treatment and orthognathic surgery. Recently, deep learning approaches have been proposed for automatic 3D cephalometric analysis based on landmarking from CBCT scans. However, these approaches have relied on uniform datasets from a single center or imaging device but without considering patient ethnicity. In addition, previous works have considered a limited number of clinically relevant cephalometric landmarks and the approaches were computationally infeasible, both impairing integration into clinical workflow. Here our aim is to analyze the clinical applicability of a light-weight deep learning neural network for fast localization of 46 clinically significant cephalometric landmarks with multi-center, multi-ethnic, and multi-device data consisting of 309 CBCT scans from Finnish and Thai patients. The localization performance of our approach resulted in the mean distance of 1.99 ± 1.55 mm for the Finnish cohort and 1.96 ± 1.25 mm for the Thai cohort. This performance turned out to be clinically significant i.e., ≤ 2 mm with 61.7% and 64.3% of the landmarks with Finnish and Thai cohorts, respectively. Furthermore, the estimated landmarks were used to measure cephalometric characteristics successfully i.e., with ≤ 2 mm or ≤ 2° error, on 85.9% of the Finnish and 74.4% of the Thai cases. Between the two patient cohorts, 33 of the landmarks and all cephalometric characteristics had no statistically significant difference (p < 0.05) measured by the Mann-Whitney U test with Benjamini-Hochberg correction. Moreover, our method is found to be computationally light, i.e., providing the predictions with the mean duration of 0.77 s and 2.27 s with single machine GPU and CPU computing, respectively. Our findings advocate for the inclusion of this method into clinical settings based on its technical feasibility and robustness across varied clinical datasets.

Surface defect detection and semantic segmentation with a novel lightweight deep neural network

Current approaches to defect detection and segmentation make essential use of machine learning methods. To develop lightweight models is one of key tasks for many defect detection and segmentation applications. In this work, we present a lightweight trilateral parallel feature extraction with multi-feature aggregation network (TriMFANet) for surface defect detection and segmentation. In TriMFANet, the top lateral is the feature-rich extraction used to capture detailed information. The other two laterals, efficient semantic feature extraction (ESFE) and reverse ESFE, leverage Hadamard product attention to jointly extract deep-level global feature information. Additionally, the MFA module employs origin-symmetric sigmoid attention to enhance deep feature information and integrates the triple features. We conducted binary defect segmentation tasks on the SD-saliency-900 and RSDDs datasets, achieving outstanding performance in both S α and E ξ . For multi-class defect detection tasks on the NEU-Seg and MSD datasets, we rank first with mIoU scores of 79.0% and 81.2% respectively. Experimental results demonstrate that our lightweight model with only 90 K parameters exhibits excellent performance.

Fine-grained food image classification and recipe extraction using a customized deep neural network and NLP

Global eating habits cause health issues leading people to mindful eating. This has directed attention to applying deep learning to food-related data. The proposed work develops a new framework integrating neural network and natural language processing for classification of food images and automated recipe extraction. It address the challenges of intra-class variability and inter-class similarity in food images that have received shallow attention in the literature. Firstly, a customized lightweight deep convolution neural network model, MResNet-50 for classifying food images is proposed. Secondly, automated ingredient processing and recipe extraction is done using natural language processing algorithms: Word2Vec and Transformers in conjunction. Thirdly, a representational semi-structured domain ontology is built to store the relationship between cuisine, food item, and ingredients. The accuracy of the proposed framework on the Food-101 and UECFOOD256 datasets is increased by 2.4% and 7.5%, respectively, outperforming existing models in literature such as DeepFood, CNN-Food, Wiser, and other pre-trained neural networks.

Evaluation of strength and stiffness degradation of RC shear walls: An integrated image processing and deep learning approach

AbstractIn the aftermath of an earthquake, damage detection and performance evaluation of structural components are imperative for assessing the residual seismic capacity of a building. In this study, an integrated image processing and deep learning approach was developed to evaluate the degradation in strength and stiffness (i.e., strength reduction and stiffness reduction) of reinforced concrete (RC) shear walls. The approach comprised two main tasks: detecting and localizing visible seismic damage from photographs and evaluating strength and stiffness degradation based on this information. The semantic segmentation network, Damage‐Net, was used for damage detection and localization. A novel crack morphological processing layer and a patch feature extraction layer were developed for damage feature extraction and compression. A lightweight deep convolutional neural network named DegradeEval‐Net_v2, featuring the upgraded dilated and separable convolution block and multi‐layer perception, was developed to link the damage feature with strength and stiffness degradation. A database comprising test data and photographs of 14 RC shear wall specimens with a flexural‐dominated behavior mode and high to intermediate ductility was constructed to train and test the DegradeEval‐Net_v2 network. The results indicate that DegradeEval‐Net_v2 substantially improved the performance assessment accuracy of damaged RC shear walls, with a 35% smaller root mean square error (RMSE) for stiffness degradation evaluation and 75% smaller RMSE for strength degradation evaluation, compared with the provisions specified in JBDPA and FEMA guidelines. Moreover, evaluation results on test sets demonstrate that introducing the damage feature extraction and compression layers effectively preserved local crack information and improved the accuracy with which stiffness reduction was evaluated. In addition, DegradeEval‐Net_v2 outperformed ResNet18 and MobileNet V3 in terms of balanced efficiency and accuracy. Interpretability analysis demonstrates that the model learned the distinct contribution patterns of various visible damage indexes to stiffness and strength degradation across different loading levels.

Lightweight Deep Neural Network with Data Redundancy Removal and Regression for DOA Estimation in Sensor Array

Lightweight Deep Neural Network with Data Redundancy Removal and Regression for DOA Estimation in Sensor Array

Explaining decisions of a light-weight deep neural network for real-time coronary artery disease classification in magnetic resonance imaging

In certain healthcare settings, such as emergency or critical care units, where quick and accurate real-time analysis and decision-making are required, the healthcare system can leverage the power of artificial intelligence (AI) models to support decision-making and prevent complications. This paper investigates the optimization of healthcare AI models based on time complexity, hyper-parameter tuning, and XAI for a classification task. The paper highlights the significance of a lightweight convolutional neural network (CNN) for analysing and classifying Magnetic Resonance Imaging (MRI) in real-time and is compared with CNN-RandomForest (CNN-RF). The role of hyper-parameter is also examined in finding optimal configurations that enhance the model’s performance while efficiently utilizing the limited computational resources. Finally, the benefits of incorporating the XAI technique (e.g. GradCAM and Layer-wise Relevance Propagation) in providing transparency and interpretable explanations of AI model predictions, fostering trust, and error/bias detection are explored. Our inference time on a MacBook laptop for 323 test images of size 100x100 is only 2.6 sec, which is merely 8 milliseconds per image while providing comparable classification accuracy with the ensemble model of CNN-RF classifiers. Using the proposed model, clinicians/cardiologists can achieve accurate and reliable results while ensuring patients’ safety and answering questions imposed by the General Data Protection Regulation (GDPR). The proposed investigative study will advance the understanding and acceptance of AI systems in connected healthcare settings.