Lightweight Deep Learning Model Research Articles

Self Distillation for Improving the Generalizability of Retinal Disease Diagnosis Using Optical Coherence Tomography Images

Optical coherence tomography (OCT) imaging has become a point-of-care imaging modality for the diagnosis of retinal diseases. Varying speckle noise in the OCT images across datasets and scanners worsens the performance of existing artificial intelligence (deep learning) models, that have been trained mostly with images having a particular noise level. The existing deep learning models for predicting retinal diseases are heavy, requires a sophisticated computing environment to train and deploy. Generalized lightweight deep learning models that can provide an automated diagnosis on an edge platform are highly appealing in the clinic. This work proposes a self distillation framework based on lightweight deep learning models for building generalizable deep models for retinal disease diagnosis. The proposed approach with three different baseline models ResNet18, MobileNetV2 and ShuffleNetV2, has been validated on simulated and real-time noisy OCT B-scans spanning a range of SNRs from four OCT datasets. The proposed method significantly outperforms the existing methods with improvement (as high as 14%) in precision, accuracy, and F1-score, to show that the self distillation framework can provide more generalizability for automated retinal diagnosis.

Catalysing assistive solutions by deploying light-weight deep learning model on edge devices

ABSTRACT Nowadays, real-time object detection, which is a crucial task, is being performed through image processing and deep learning techniques. As there are several high-performance computing edge devices available, selecting the best-fit device for a particular problem is a tough task and keeping in mind the cost, performance, and weight of the device in mind. One faces several challenges while performing this task in real-time such as a lack of resources in terms of power and mobility. We have provided an insight into the computation power of devices in terms of Frames per Second (FPS) by deploying object detection models on them. This paper will provide insight into selecting the appropriate combination of device and object detection models for real-time applications. Raspberry Pi 3 (RPi3), Raspberry Pi 4 (RPi4), Intel Neural Compute Stick 2 (NCS2), and Nvidia Jetson NANO are popular devices with high computation power used for real-time applications. The memory constraints of devices along with the deployment of different You Only Look Once (YOLO) and Single-Shot Detector (SSD) are the two object detection models that have been explained in this paper. A deep learning inference optimiser, TensorRT, has been used in NANO to achieve high throughput in the performance of object detection. The precision, recall, and F1 score achieved on deploying each tested model have been presented. After observing the devices during experimentation, RPi4+NCS2 showed the best execution with the blend of factors i.e. speed, portability, and user-friendliness.

DDD TinyML: A TinyML-Based Driver Drowsiness Detection Model Using Deep Learning.

Driver drowsiness is one of the main causes of traffic accidents today. In recent years, driver drowsiness detection has suffered from issues integrating deep learning (DL) with Internet-of-things (IoT) devices due to the limited resources of IoT devices, which pose a challenge to fulfilling DL models that demand large storage and computation. Thus, there are challenges to meeting the requirements of real-time driver drowsiness detection applications that need short latency and lightweight computation. To this end, we applied Tiny Machine Learning (TinyML) to a driver drowsiness detection case study. In this paper, we first present an overview of TinyML. After conducting some preliminary experiments, we proposed five lightweight DL models that can be deployed on a microcontroller. We applied three DL models: SqueezeNet, AlexNet, and CNN. In addition, we adopted two pretrained models (MobileNet-V2 and MobileNet-V3) to find the best model in terms of size and accuracy results. After that, we applied the optimization methods to DL models using quantization. Three quantization methods were applied: quantization-aware training (QAT), full-integer quantization (FIQ), and dynamic range quantization (DRQ). The obtained results in terms of the model size show that the CNN model achieved the smallest size of 0.05 MB using the DRQ method, followed by SqueezeNet, AlexNet MobileNet-V3, and MobileNet-V2, with 0.141 MB, 0.58 MB, 1.16 MB, and 1.55 MB, respectively. The result after applying the optimization method was 0.9964 accuracy using DRQ in the MobileNet-V2 model, which outperformed the other models, followed by the SqueezeNet and AlexNet models, with 0.9951 and 0.9924 accuracies, respectively, using DRQ.

Revolutionizing power line inspection: automated data acquisition through autonomous UAVs in simulated environment

Due to the growing need for electricity, the effective inspection of the power lines is becoming an important matter. In this paper, the author presents the inspection of power or transmission line with autonomous automatic UAVs (Unmanned Aerial Vehicles). For the comprehensive inspection of power lines and its different components, such as (cross arms, cracks in poles, rot damage, and insulator burn), It is needed to inspect from every side of the elements and the masts. So, the angle and speed of the drone are much more important to take images while moving around the poles. The simulator used for the experiments, including deep learning models, acts as a vital source of data analysis. At the same time, pictures are used as the primary data source. Through the Deep learning method, a suggestion of action generated for the movement around the masts. The use of a simulator is a quick, accurate, and inexpensive solution, with less real/world factors affecting the inspection process, such as weather, time, and cost of using many different resources. This study presents experiments with lightweight deep learning models through developing the prototype of vision based unmanned aerial vehicle to inspect the power line in a simulated environment. It focuses on the large demand of power companies to inspect the power line autonomously with the influence of deep learning. Finally, several deep learning models are compared when inspection along the power lines. The model shows satisfactory results in the testing path. The model trained by MobileNetV2 performs best among all other models.

A lightweight deep learning model based recommender system by sentiment analysis

Recommender systems based on sentiment analysis become challenging due to the presence of enormous data available over the internet. With the lack of proper data cleaning and analysis methods, existing machine learning (ML) techniques fail to generate accurate recommendations. To overcome this issue, this paper proposes a Light Deep Learning (LightDL)-based recommender system that uses Twitter-based reviews. First, the data is collected from Twitter and cleaned by subsequent data cleaning processes. Then, this pre-processed data is fed into the LightDL model, which learns the important features like hashtags, unigrams, multigrams, etc. from each piece of data. Here, we have learned about four groups of features, including semantic features, syntactic features, symbolic features, and tweet-based features. Finally, the data is classified into positive, negative, and neutral categories according to the learned features. On the basis of classified sentiment, the review is generated to the users. Finally, the model is evaluated in terms of accuracy, precision, recall, f-measure, and error rate through extensive experiments in Matlab. The proposed LightDL model outperforms in all performance measures; specifically, it achieves 95% accuracy for the Twitter dataset.

A Wireless Sensor System for Diabetic Retinopathy Grading Using MobileViT-Plus and ResNet-Based Hybrid Deep Learning Framework

Traditional fundus image-based diabetic retinopathy (DR) grading depends on the examiner’s experience, requiring manual annotations on the fundus image and also being time-consuming. Wireless sensor networks (WSNs) combined with artificial intelligence (AI) technology can provide automatic decision-making for DR grading application. However, the diagnostic accuracy of the AI model is one of challenges that limited the effectiveness of the WSNs-aided DR grading application. Regarding this issue, we propose a WSN architecture and a parallel deep learning framework (HybridLG) for actualizing automatic DR grading and achieving a fundus image-based deep learning model with superior classification performance, respectively. In particular, the framework constructs a convolutional neural network (CNN) backbone and a Transformer backbone in a parallel manner. A novel lightweight deep learning model named MobileViT-Plus is proposed to implement the Transformer backbone of the HybridLG, and a model training strategy inspired by an ensemble learning strategy is designed to improve the model generalization ability. Experimental results demonstrate the state-of-the-art performance of the proposed HybridLG framework, obtaining excellent performance in grading diabetic retinopathy with strong generalization performance. Our work is significant for guiding the studies of WSNs-aided DR grading and providing evidence for supporting the efficacy of the AI technology in DR grading applications.

AFF-YOLOX: An improved lightweight YOLOX network to detect early hatching information of duck eggs

Early hatching information detection of duck eggs is a crucial step in the duck egg hatching industry. In this study, we proposed a lightweight deep learning model AFF-YOLOX to detect the hatching information of multiple duck eggs on the standard hatching trays, meeting the speed requirements of real production. Based on YOLOX of networks structure, we added the AFF to the PA-FPN. Additionally, the input size of the image was changed according to the hatching tray and remove the redundant part of the network. The results showed that AFF-YOLOX could significantly improve the detection effect compared with the original YOLOX. The mAP was improved from 74.30% to 97.55%, the number of parameters was reduced from 99 M to 0.49 M, and the latency per image was decreased from 556 ms to 178 ms. AFF-YOLOX can detect abnormal eggs on incubation day 4 and can completely identify all abnormal eggs on incubation day 6. The results demonstrated that the optimization and lightweight methods we are effective and provides a technical reference for automatic detection of hatching information of duck eggs in the hatchery industry.

Design of a Sleep Modulation System with FPGA-Accelerated Deep Learning for Closed-loop Stage-Specific In-Phase Auditory Stimulation.

Closed-loop sleep modulation is an emerging research paradigm to treat sleep disorders and enhance sleep benefits. However, two major barriers hinder the widespread application of this research paradigm. First, subjects often need to be wire-connected to rack-mount instrumentation for data acquisition, which negatively affects sleep quality. Second, conventional real-time sleep stage classification algorithms give limited performance. In this work, we conquer these two limitations by developing a sleep modulation system that supports closed-loop operations on the device. Sleep stage classification is performed using a lightweight deep learning (DL) model accelerated by a low-power field-programmable gate array (FPGA) device. The DL model uses a single channel electroencephalogram (EEG) as input. Two convolutional neural networks (CNNs) are used to capture general and detailed features, and a bidirectional long-short-term memory (LSTM) network is used to capture time-variant sequence features. An 8-bit quantization is used to reduce the computational cost without compromising performance. The DL model has been validated using a public sleep database containing 81 subjects, achieving a state-of-the-art classification accuracy of 85.8% and a F1-score of 79%. The developed model has also shown the potential to be generalized to different channels and input data lengths. Closed-loop in-phase auditory stimulation has been demonstrated on the test bench.

Lightweight Deep Learning Model for Radar-Based Fall Detection With Metric Learning

The radar-based fall detection system has grown in popularity because of its stability and privacy protection. Deep neural networks have been used in previous radar-based fall detection systems to improve detection accuracy. However, most of them need a lot of memory and have high computational complexity, making them impractical for Internet of Things (IoT) devices. In this paper, we propose an extremely lightweight network named Tiny-RadarNet for extracting characteristics from raw data. Unlike traditional neural networks, we use a unique parallel one-dimensional depthwise convolutions structure as the core module to eliminate the need for standard convolutions, and achieve significant parameter reduction. Furthermore, instead of regarding fall detection as a classification problem, we use our metric learning technique to treat it as a matching problem for better distinction of embeddings. Finally, we introduce a novel dual loss function to improve the proposed network’s robustness against unobserved human motions without the need for additional anchors or expensive computation. The experimental results reveal that the proposed method can achieve comparable fall detection accuracy compared with state-of-art methods but with much fewer weight parameters and lower computational complexity. These findings suggest that a low-power and low-latency fall detection solution for IoT applications is achievable.

Comparative Study on Distributed Lightweight Deep Learning Models for Road Pothole Detection.

This paper delves into image detection based on distributed deep-learning techniques for intelligent traffic systems or self-driving cars. The accuracy and precision of neural networks deployed on edge devices (e.g., CCTV (closed-circuit television) for road surveillance) with small datasets may be compromised, leading to the misjudgment of targets. To address this challenge, TensorFlow and PyTorch were used to initialize various distributed model parallel and data parallel techniques. Despite the success of these techniques, communication constraints were observed along with certain speed issues. As a result, a hybrid pipeline was proposed, combining both dataset and model distribution through an all-reduced algorithm and NVlinks to prevent miscommunication among gradients. The proposed approach was tested on both an edge cluster and Google cluster environment, demonstrating superior performance compared to other test settings, with the quality of the bounding box detection system meeting expectations with increased reliability. Performance metrics, including total training time, images/second, cross-entropy loss, and total loss against the number of the epoch, were evaluated, revealing a robust competition between TensorFlow and PyTorch. The PyTorch environment's hybrid pipeline outperformed other test settings.

Artificial Intelligence for Media Ecological Integration and Knowledge Management

Information Technology’s development increases day by day, making life easier in terms of work and progress. In these developments, knowledge management is becoming mandatory in all the developing sectors. However, the conventional model for growth analysis in organizations is tedious as data are maintained in ledgers, making the process time consuming. Media Ecology, a new trending technology, overcomes this drawback by being integrated with artificial intelligence. Various sectors implement this integrated technology. The marketing strategy of Huawei Technologies Co. Ltd. is analyzed in this research to examine the advantages of Media Ecology Technology in integration with artificial intelligence and a Knowledge Management Model. This combined model supports sensor technology by considering each medium, the data processing zone, and user location as nodes. A Q-R hybrid simulation methodology is implemented to analyze the data collected through Media Ecology. The proposed method is compared with the inventory model, and the results show that the proposed system provides increased profit to the organization. Paying complete attention to Artificial intelligence without the help of lightweight deep learning models is impossible. Thus, lightweight deep models have been introduced in most situations, such as healthcare management, maintenance systems, and controlling a few IoT devices. With the support of high-power consumption as computational energy, it adapts to lightweight devices such as mobile phones. One common expectation from the deep learning concept is to develop an optimal structure in case time management.

Lightweight Deep Learning Model for Real-Time Colorectal Polyp Segmentation

In colonoscopy, computer vision and artificial intelligence technology have enabled the automatic detection of the location of polyps and their visualization. These advancements have facilitated considerable research in the field. However, deep learning models used in the segmentation problem for capturing various patterns of polyps are becoming increasingly complex, which has rendered their operation in real time difficult. To identify and overcome this problem, a study was conducted on a model capable of precise polyp segmentation while increasing its processing speed. First, an efficient, high-performance, and lightweight model suitable for the segmentation of polyps was sought; the performance of existing segmentation models was compared and combined to obtain a learning model that exhibited good accuracy and speed. Next, hyperparameters were found for the MobileNetV3-encoder-based DeepLabV3+ model and, after tuning the hyperparameters, quantitative and qualitative results were compared, and the final model was selected. The experimental results showed that this model achieved high accuracy, with a Dice coefficient of 93.79%, while using a limited number of parameters and computational resources. Specifically, the model used 6.18 million parameters and 1.623 giga floating point operations for the CVC-ClinicDB dataset. This study revealed that increasing the amount of computation and parameters of the model did not guarantee unconditional performance. Furthermore, for the search and removal of polyps in cases in which morphological information is critical, an efficient model with low model complexity and high accuracy was proposed for real-time segmentation.

A Lightweight Deep Learning Model for Automatic Modulation Classification Using Residual Learning and Squeeze–Excitation Blocks

Automatic modulation classification (AMC) is a vital process in wireless communication systems that is fundamentally a classification problem. It is employed to automatically determine the type of modulation of a received signal. Deep learning (DL) methods have gained popularity in addressing the problem of modulation classification, as they automatically learn the features without needing technical expertise. However, their efficacy depends on the complexity of the algorithm, which can be characterized by the number of parameters. In this research, we presented a deep learning algorithm for AMC, inspired by residual learning, which has remarkable accuracy and great representational ability. We also employed a squeeze-and-excitation network that is capable of exploiting modeling interconnections between channels and adaptively re-calibrates the channel-wise feature response to improve performance. The proposed network was designed to meet the accuracy requirements with a reduced number of parameters for efficiency. The proposed model was evaluated on two benchmark datasets and compared with existing methods. The results show that the proposed model outperforms existing methods in terms of accuracy and has up to 72.5% fewer parameters than convolutional neural network designs.

Image Forgery Detection based on Fusion of Lightweight Deep Learning Models

Image Forgery Detection based on Fusion of Lightweight Deep Learning Models

Lightweight Model for Botnet Attack Detection in Software Defined Network-Orchestrated IoT

The Internet of things (IoT) is being used in a variety of industries, including agriculture, the military, smart cities and smart grids, and personalized health care. It is also being used to control critical infrastructure. Nevertheless, because the IoT lacks security procedures and lack the processing power to execute computationally costly antimalware apps, they are susceptible to malware attacks. In addition, the conventional method by which malware-detection mechanisms identify a threat is through known malware fingerprints stored in their database. However, with the ever-evolving and drastic increase in malware threats in the IoT, it is not enough to have traditional antimalware software in place, which solely defends against known threats. Consequently, in this paper, a lightweight deep learning model for an SDN-enabled IoT framework that leverages the underlying IoT resource-constrained devices by provisioning computing resources to deploy instant protection against botnet malware attacks is proposed. The proposed model can achieve 99% precision, recall, and F1 score and 99.4% accuracy. The execution time of the model is 0.108 milliseconds with 118 KB size and 19,414 parameters. The proposed model can achieve performance with high accuracy while utilizing fewer computational resources and addressing resource-limitation issues.

Modeling air quality PM2.5 forecasting using deep sparse attention-based transformer networks

Air quality forecasting is of great importance in environmental protection, government decision-making, people's daily health, etc. Existing research methods have failed to effectively modeling long-term and complex relationships in time series PM2.5 data and exhibited low precision in long-term prediction. To address this issue, in this paper a new lightweight deep learning model using sparse attention-based Transformer networks (STN) consisting of encoder and decoder layers, in which a multi-head sparse attention mechanism is adopted to reduce the time complexity, is proposed to learn long-term dependencies and complex relationships from time series PM2.5 data for modeling air quality forecasting. Extensive experiments on two real-world datasets in China, i.e., Beijing PM2.5 dataset and Taizhou PM2.5 dataset, show that our proposed method not only has relatively small time complexity, but also outperforms state-of-the-art methods, demonstrating the effectiveness of the proposed STN method on both short-term and long-term air quality prediction tasks. In particular, on singe-step PM2.5 forecasting tasks our proposed method achieves R2 of 0.937 and reduces RMSE to 19.04 µg/m3 and MAE to 11.13 µg/m3 on Beijing PM2.5 dataset. Also, our proposed method obtains R2 of 0.924 and reduces RMSE to 5.79 µg/m3 and MAE to 3.76 µg/m3 on Taizhou PM2.5 dataset. For long-term time step prediction, our proposed method still performs best among all used methods on multi-step PM2.5 forecasting results for the next 6, 12, 24, and 48 h on two real-world datasets.

LightSeizureNet: A Lightweight Deep Learning Model for Real-Time Epileptic Seizure Detection.

The monitoring of epilepsy patients in non-hospital environment is highly desirable, where ultra-low power wearable seizure detection devices are essential in such a system. The state-of-the-art epileptic seizure detection algorithms targeting such devices either rely on manual feature extractions, which can be biased due to the experience of experts, or deep neural networks, which suffer from high computation complexity. In this paper, we propose a lightweight deep learning model, LightSeizureNet (LSN), for real-time epileptic seizure detection based on raw EEG data in ultra-low power wearable seizure detection devices. The proposed LSN model includes a patient-independent version and a patient-specific version, both of which avoids manual feature extractions and high computation complexity, while maintaining good classification accuracy. Dilated one-dimensional (1D) convolution, global average pooling, and kernel-wise pruning are adopted to compress the LSN model. The proposed models are evaluated on the CHB-MIT scalp EEG database. The patient-independent LSN model achieves 97.09% accuracy with 6.2M MACs, while the patient-specific LSN model achieves 99.77% accuracy with 3.7 M MACs, which are competitive compared to the state of the art in terms of accuracy and complexity. Furthermore, the proposed model is highly interpretable, which is missing in many previous works. By using a uniform approach to explore the interpretability of the proposed model, fine-grained information such as the activated brain region and the frequency of brainwave during seizures is obtained for clinical diagnosis.

Accurate Detection of Alzheimer's Disease Using Lightweight Deep Learning Model on MRI Data.

Alzheimer's disease (AD) is a neurodegenerative disorder characterized by cognitive impairment and aberrant protein deposition in the brain. Therefore, the early detection of AD is crucial for the development of effective treatments and interventions, as the disease is more responsive to treatment in its early stages. It is worth mentioning that deep learning techniques have been successfully applied in recent years to a wide range of medical imaging tasks, including the detection of AD. These techniques have the ability to automatically learn and extract features from large datasets, making them well suited for the analysis of complex medical images. In this paper, we propose an improved lightweight deep learning model for the accurate detection of AD from magnetic resonance imaging (MRI) images. Our proposed model achieves high detection performance without the need for deeper layers and eliminates the use of traditional methods such as feature extraction and classification by combining them all into one stage. Furthermore, our proposed method consists of only seven layers, making the system less complex than other previous deep models and less time-consuming to process. We evaluate our proposed model using a publicly available Kaggle dataset, which contains a large number of records in a small dataset size of only 36 Megabytes. Our model achieved an overall accuracy of 99.22% for binary classification and 95.93% for multi-classification tasks, which outperformed other previous models. Our study is the first to combine all methods used in the publicly available Kaggle dataset for AD detection, enabling researchers to work on a dataset with new challenges. Our findings show the effectiveness of our lightweight deep learning framework to achieve high accuracy in the classification of AD.

Designing Efficient and Lightweight Deep Learning Models for Healthcare Analysis

Designing Efficient and Lightweight Deep Learning Models for Healthcare Analysis

YOLO-Based Light-Weight Deep Learning Models for Insect Detection System with Field Adaption

The most incredible diversity, abundance, spread, and adaptability in biology are found in insects. The foundation of insect study and pest management is insect recognition. However, most of the current insect recognition research depends on a small number of insect taxonomic experts. We can use computers to differentiate insects accurately instead of professionals because of the quick advancement of computer technology. The “YOLOv5” model, with five different state of the art object detection techniques, has been used in this insect recognition and classification investigation to identify insects with the subtle differences between subcategories. To enhance the critical information in the feature map and weaken the supporting information, both channel and spatial attention modules are introduced, improving the network’s capacity for recognition. The experimental findings show that the F1 score approaches 0.90, and the mAP value reaches 93% through learning on the self-made pest dataset. The F1 score increased by 0.02, and the map increased by 1% as compared to other YOLOv5 models, demonstrating the success of the upgraded YOLOv5-based insect detection system.