Lightweight Model Research Articles

SSBlazer: a genome-wide nucleotide-resolution model for predicting single-strand break sites

Single-strand breaks are the major DNA damage in the genome and serve a crucial role in various biological processes. To reveal the significance of single-strand breaks, multiple sequencing-based single-strand break detection methods have been developed, which are costly and unfeasible for large-scale analysis. Hence, we propose SSBlazer, an explainable and scalable deep learning framework for single-strand break site prediction at the nucleotide level. SSBlazer is a lightweight model with robust generalization capabilities across various species and is capable of numerous unexplored SSB-related applications.

A Weakly Supervised Hybrid Lightweight Network for Efficient Crowd Counting

Crowd-counting networks have become the mainstream method to deploy crowd-counting techniques on resource-constrained devices. Significant progress has been made in this field, with many outstanding lightweight models being proposed successively. However, challenges like scare-variation, global feature extraction, and fine-grained head annotation requirements still exist in relevant tasks, necessitating further improvement. In this article, we propose a weakly supervised hybrid lightweight crowd-counting network that integrates the initial layers of GhostNet as the backbone to efficiently extract local features and enrich intermediate features. The incorporation of a modified Swin-Transformer block addresses the need for effective global context information. A Pyramid Pooling Aggregation Module handles the inherent scale variation problem in crowd-counting tasks in a more computation-efficient way. This module, along with the cross-attention module, serves as bridges to promote the feature information flow between local features and global context information. Finally, a simplified regressor module is designed to allow the proposed model with weakly supervised guidance for training to avoid precise location-level annotations, noting that the omission of density map generation makes the proposed network more lightweight. Our results on the UCF-QNRF dataset indicate our model is 8.73% and 12.17% more accurate on MAE and MSE metrics, respectively, than the second-best ARNet, in which the parameters decrease by 4.52%. On the ShanghaiTech A dataset, MAE and MSE drop 1.5% and 3.2%, respectively, compared to the second-best PDDNet. The experimental results for accuracy and inference speed evaluation on some mainstream datasets validate the effective design principle of our model.

Machine Learning-based Intrusion Detection Technique for IoT: Simulation with Cooja

The Internet of Things (IoT) is one of the promising technologies of the future. It offers many attractive features that we depend on nowadays with less effort and faster in real-time. However, it is still vulnerable to various threats and attacks due to the obstacles of its heterogeneous ecosystem, adaptive protocols, and self-configurations. In this paper, three different 6LoWPAN attacks are implemented in the IoT via Contiki OS to generate the proposed dataset that reflects the 6LoWPAN features in IoT. For analyzed attacks, six scenarios have been implemented. Three of these are free of malicious nodes, and the others scenarios include malicious nodes. The typical scenarios are a benchmark for the malicious scenarios for comparison, extraction, and exploration of the features that are affected by attackers. These features are used as criteria input to train and test our proposed hybrid Intrusion Detection and Prevention System (IDPS) to detect and prevent 6LoWPAN attacks in the IoT ecosystem. The proposed hybrid IDPS has been trained and tested with improved accuracy on both KoU-6LoWPAN-IoT and Edge IIoT datasets. In the proposed hybrid IDPS for the detention phase, the Artificial Neural Network (ANN) classifier achieved the highest accuracy among the models in both the 2-class and N-class. Before the accuracy improved in our proposed dataset with the 4-class and 2-class mode, the ANN classifier achieved 95.65% and 99.95%, respectively, while after the accuracy optimization reached 99.84% and 99.97%, respectively. For the Edge IIoT dataset, before the accuracy improved with the 15-class and 2-class modes, the ANN classifier achieved 95.14% and 99.86%, respectively, while after the accuracy optimized up to 97.64% and 99.94%, respectively. Also, the decision tree-based models achieved lightweight models due to their lower computational complexity, so these have an appropriate edge computing deployment. Whereas other ML models reach heavyweight models and are required more computational complexity, these models have an appropriate deployment in cloud or fog computing in IoT networks.

A Lightweight Algorithm for Recognizing Pear Leaf Diseases in Natural Scenes Based on an Improved YOLOv5 Deep Learning Model

The precise detection of diseases is crucial for the effective treatment of pear trees and to improve their fruit yield and quality. Currently, recognizing plant diseases in complex backgrounds remains a significant challenge. Therefore, a lightweight CCG-YOLOv5n model was designed to efficiently recognize pear leaf diseases in complex backgrounds. The CCG-YOLOv5n model integrates a CA attention mechanism, CARAFE up-sampling operator, and GSConv into YOLOv5n. It was trained and validated using a self-constructed dataset of pear leaf diseases. The model size and FLOPs are only 3.49 M and 3.8 G, respectively. The mAP@0.5 is 92.4%, and the FPS is up to 129. Compared to other lightweight indicates that the models, the experimental results demonstrate that the CCG-YOLOv5n achieves higher average detection accuracy and faster detection speed with a smaller computation and model size. In addition, the robustness comparison test CCG-YOLOv5n model has strong robustness under various lighting and weather conditions, including frontlight, backlight, sidelight, tree shade, and rain. This study proposed a CCG-YOLOv5n model for accurately detecting pear leaf diseases in complex backgrounds. The model is suitable for use on mobile terminals or devices.

IPMCNet: A Lightweight Algorithm for Invasive Plant Multiclassification

Invasive plant species pose significant biodiversity and ecosystem threats. Real-time identification of invasive plants is a crucial prerequisite for early and timely prevention. While deep learning has shown promising results in plant recognition, the use of deep learning models often involve a large number of parameters and high data requirements for training. Unfortunately, the available data for various invasive plant species are often limited. To address this challenge, this study proposes a lightweight deep learning model called IPMCNet for the identification of multiple invasive plant species. IPMCNet attains high recognition accuracy even with limited data and exhibits strong generalizability. Simultaneously, by employing depth-wise separable convolutional kernels, splitting channels, and eliminating fully connected layer, the model’s parameter count is lower than that of some existing lightweight models. Additionally, the study explores the impact of different loss functions, and the insertion of various attention modules on the model’s accuracy. The experimental results reveal that, compared with eight other existing neural network models, IPMCNet achieves the highest classification accuracy of 94.52%. Furthermore, the findings suggest that focal loss is the most effective loss function. The performance of the six attention modules is suboptimal, and their insertion leads to a decrease in model accuracy.

Enhancing Emotion Recognition through Federated Learning: A Multimodal Approach with Convolutional Neural Networks

Human–machine interaction covers a range of applications in which machines should understand humans’ commands and predict their behavior. Humans commonly change their mood over time, which affects the way we interact, particularly by changing speech style and facial expressions. As interaction requires quick decisions, low latency is critical for real-time processing. Edge devices, strategically placed near the data source, minimize processing time, enabling real-time decision-making. Edge computing allows us to process data locally, thus reducing the need to send sensitive information further through the network. Despite the wide adoption of audio-only, video-only, and multimodal emotion recognition systems, there is a research gap in terms of analyzing lightweight models and solving privacy challenges to improve model performance. This motivated us to develop a privacy-preserving, lightweight, CNN-based (CNNs are frequently used for processing audio and video modalities) audiovisual emotion recognition model, deployable on constrained edge devices. The model is further paired with a federated learning protocol to preserve the privacy of local clients on edge devices and improve detection accuracy. The results show that the adoption of federated learning improved classification accuracy by ~2%, as well as that the proposed federated learning-based model provides competitive performance compared to other baseline audiovisual emotion recognition models.

Predicting optimal sparse general matrix-matrix multiplication algorithm on GPUs

Sparse General Matrix-Matrix Multiplication (SpGEMM) has played an important role in a number of applications. So far, many efficient algorithms have been proposed to improve the performance of SpGEMM on GPUs. However, the performance of each algorithm for matrices of different structures varies a lot. There is no algorithm that can achieve the optimal performance of SpGEMM computation on all matrices. In this article, we design a machine learning based approach for predicting the optimal SpGEMM algorithm on input matrices. By extracting features from input matrices, we utilize LightGBM and XGBoost to train different lightweight models. The models are capable of predicting the best performing algorithm with low inference overhead and high accuracy for the given input matrices. We also investigate the impact of tree depth on model accuracy and inference overhead. Our evaluation shows that an increase in tree depth leads to a corresponding increase in prediction accuracy, reaching a maximum of approximately 85%, while resulting in increased inference overhead of approximately 10 µs. Compared with the state-of-the-art algorithms on three GPU platforms, our method achieves better overall performance.

Improvement of lithium battery corner detection accuracy based on image restoration method

Target detection technology has been widely used in the automatic production of lithium batteries. However, motion blur will lead to the reduction of the angular position detection accuracy of lithium batteries. To solve this problem, an improved fuzzy recovery model for angular position of lithium battery is proposed in this paper. Firstly, the improved lightweight neural network RepVGG was used as the main module of the backbone network, so that the network could improve the performance of network feature extraction while reducing the number of calculation parameters and improving the reasoning speed of fuzzy restoration. Secondly, we optimize the multi-Dconv head transposed attention (MDTA) module and reference it to the generator, which reduces the complexity of the model and strengthens the network’s attention to details and textures, and improves the visual effect of the restored image. Finally, we design a lightweight globally connectable residual network called SAC Block and use it to to improve the discriminator, which enhances the global receptive field of the model and improves the structural similarity between the restored image and the original image. In order to verify the effectiveness of the method, we verify it on the self-built dataset and GoPro dataset. The experiments show that our proposed lightweight model improves the peak signal-to-noise ratio (PSNR) index by 9.2% and 8.6% respectively compared with the original model. The visual effect of the restored image is better than that of other current similar algorithms, and it is confirmed that our model can better improve the accuracy of lithium battery angular position detection.

Accurate and efficient galaxy classification based on mobile vision transformer

Understanding the formation and evolution of galaxies in observational cosmology heavily relies on galaxy morphological classification. Nevertheless, the continuously growing volume of astronomical data has surpassed human capacity for manual classification. In this context, deep learning presents a promising approach to enhancing classifying galaxies. In this paper, the Mobile Vision Transformer (MobileViT) is introduced to construct an efficient and accurate galaxy classifier. Transfer learning is introduced to assist in model fine-tuning. MobileViT combines the features of MobileNet and Visual Transformer (ViT). A lightweight model is used to effectively analyse the relationships between sequences for efficient and accurate classification. Experiments are built on Galaxy10 DECals dataset. Excellent performance is achieved in identifying galaxy types compared to other lightweight models. The model achieves an accuracy of over 87% and maintains a high speed of inference of less than 50 milliseconds per step. Experimental results show that the introduction of MobileViT is the best solution for efficient galaxy classification. The model can be deployed on any portable device for instant observation and classification.

LiteMixer: Cauliflower Disease Diagnosis based on a Novel Lightweight Neural Network

Abstract Cauliflower, a globally cultivated and nutritionally rich crop, confronts significant challenges in quality and yield due to the rising prevalence of diseases. Traditional manual detection methods, suitable for empiricists or plant pathologists, prove inefficient. Furthermore, existing automated disease identification methods in cauliflower often neglect crucial computational performance metrics within computer vision algorithms, such as complexity, inference speed and training time. This study introduces LiteMixer, a novel lightweight model designed to address these challenges. The Lightweight Mixed-Domain Feature Extraction module (LMFE) meticulously captures global image features, followed by a maximum pooling layer that downscales the resulting multidimensional feature matrix. The Plug-and-Play Multi-Scale Lightweight Convolutional Attention Fusion module (MLCAF) integrates multichannel spatial features, connecting to fully connected layers for the final classification. Ablation experiments highlight the effectiveness of the LMFE module coupled with the MLCAF module. Comparative analyses against state-of-the-art and other lightweight models demonstrate LiteMixer achieving the highest accuracy in identifying cauliflower diseases at 99.86%. Notably, LiteMixer exhibits optimal computational performance, featuring minimal storage costs (4.02M) and the lowest parameter count, resulting in cost-effective computational expenses (16.78M). LiteMixer also boasts the fastest inference time (4.69 ms) and the shortest training time (865 s). This study positions LiteMixer as an advanced solution for diagnosing cauliflower leaf diseases in agricultural settings, underscoring its efficacy and practicality in overcoming the unique challenges associated with cauliflower disease detection within the realm of computer vision algorithms.

Collaborative Human Recognition With Lightweight Models in Drone-Based Search and Rescue Operations

Collaborative Human Recognition With Lightweight Models in Drone-Based Search and Rescue Operations

Numerical solutions to linear transfer problems of polarized radiation

Context. A relevant class of radiative transfer problems for polarized radiation is linear, or can be linearized, and can thus be reframed as linear systems once discretized. In this context, depending on the considered physical models, there are both highly coupled and computationally expensive problems, for which state-of-the-art iterative methods struggle to converge, and lightweight ones, for which solutions can be obtained efficiently. Aims. This work aims to exploit lightweight physical models as preconditioners for iterative solution strategies to obtain accurate and fast solutions for more complex problems. Methods. We considered a highly coupled linear transfer problem for polarized radiation, which we solved iteratively using a matrix-free generalized minimal residual (GMRES) method. Different preconditioners and initial guesses, designed in a physics-based framework, are proposed and analyzed. The action of preconditioners was also computed by applying GMRES. The overall approach thus consists of two nested GMRES iterations, one for the original problem and one for its lightweight version. As a benchmark, we considered the modeling of the intensity and polarization of the solar Ca I 4227 Å line, the Sr II 4077 Å line, and the Mg II h&k lines in a semi-empirical 1D atmospheric model, accounting for partial frequency redistribution effects in scattering processes and considering a general angle-dependent treatment. Results. Numerical experiments show that using tailored preconditioners based on simplified models of the considered problem has a noticeable impact, reducing the number of iterations to convergence by a factor of 10–20. Conclusions. By designing efficient preconditioners in a physics-based context, it is possible to significantly improve the convergence of iterative processes, obtaining fast and accurate numerical solutions to the considered problems. The presented approach is general, requiring only the selection of an appropriate lightweight model, and can be applied to a larger class of radiative transfer problems in combination with arbitrary iterative procedures.

Light-Weight Deep Learning Model for Accelerating the Classification of Mango-Leaf Disease

Mango leaf diseases represent a serious threat to world agriculture, necessitating prompt and accurate detection to avert catastrophic effects. In response, this study suggests a light-weight, deep learning-based method for automatically classifying mango leaf diseases. The model is based on the original DenseNet architecture, which is well known for its effectiveness in image classification tasks. Custom layers have been added over the existing layer of the original DenseNet model. The proposed model has been compared with other existing pre-trained models. Based on comparisons, the proposed model, DenseNet78, proved to be efficient even on a relatively small dataset, where the conventional model failed. The proposed model ensured generalization across regions, disease variants, and diverse datasets of mango leaves. The results demonstrate that the fine-tuned DenseNet architecture (DenseNet78), along with an ideal growth rate, modifying block size, and a number of layers, provides optimum accuracy, with 99.47% accuracy in identifying healthy mango leaves and 99.44% accuracy in detecting various mango leaf diseases. The results also demonstrate that the model is effective in accelerating the training process because of careful comparative analysis of all the available alternatives, including the most effective combination of optimizers, learning rate schedulers, and loss functions. The study's conclusion is an automated approach for diagnosing mango leaf disease using an improved and optimized DenseNet architecture (DenseNet78). Doi: 10.28991/ESJ-2024-08-01-03 Full Text: PDF

Multitargets Joint Training Lightweight Model for Object Detection of Substation.

The object detection of the substation is the key to ensuring the safety and reliable operation of the substation. The traditional image detection algorithms use the corresponding texture features of single-class objects and would not handle other different class objects easily. The object detection algorithm based on deep networks has generalization, and its sizeable complex backbone limits the application in the substation monitoring terminals with weak computing power. This article proposes a multitargets joint training lightweight model. The proposed model uses the feature maps of the complex model and the labels of objects in images as training multitargets. The feature maps have deeper feature information, and the feature maps of complex networks have higher information entropy than lightweight networks have. This article proposes the heat pixels method to improve the adequate object information because of the imbalance of the proportion between the foreground and the background. The heat pixels method is designed as a kind of reverse network calculation and reflects the object's position to the pixels of the feature maps. The temperature of the pixels indicates the probability of the existence of the objects in the locations. Three different lightweight networks use the complex model feature maps and the traditional tags as the training multitargets. The public dataset VOC and the substation equipment dataset are adopted in the experiments. The experimental results demonstrate that the proposed model can effectively improve object detection accuracy and reduce the time-consuming and calculation amount.

A NEW APPROACH FOR SPEECH EMOTION RECOGNITION USING SINGLE LAYERED CONVOLUTIONAL NEURAL NETWORK

Creating a computational device to identify human emotions via voice analysis represents a notable achievement in the sector of human-computer interaction, especially within the healthcare domain. We propose a new light-weight model for addressing challenges of emotions recognition. The model works based on CNN with change of kernel processing. The proposed model performs a direct matching to recognize speech emotions of different eight categories using a statistical model named Analysis of Variance (ANOVA) as kernel for features extraction and Cosine Similarity Measurement (CSM) as activation function for CNN model. This proposed model contains eight-folded single-layered intermediate neurons, and each neuron can segregate speech emotion pattern using CSM from the voice convergence matrix to explore a part of the solution from the whole solution. Experiment results demonstrates that the proposed model outperforms compared with multiple layered existing CNN methods in identifying the emotional state of a speaker.

A NEW APPROACH FOR SPEECH EMOTION RECOGNITION USING SINGLE LAYERED CONVOLUTIONAL NEURAL NETWORK

Creating a computational device to identify human emotions via voice analysis represents a notable achievement in the sector of human-computer interaction, especially within the healthcare domain. We propose a new light-weight model for addressing challenges of emotions recognition. The model works based on CNN with change of kernel processing. The proposed model performs a direct matching to recognize speech emotions of different eight categories using a statistical model named Analysis of Variance (ANOVA) as kernel for features extraction and Cosine Similarity Measurement (CSM) as activation function for CNN model. This proposed model contains eight-folded single-layered intermediate neurons, and each neuron can segregate speech emotion pattern using CSM from the voice convergence matrix to explore a part of the solution from the whole solution. Experiment results demonstrates that the proposed model outperforms compared with multiple layered existing CNN methods in identifying the emotional state of a speaker.

YOLOSeaShip: a lightweight model for real-time ship detection

ABSTRACT With the rapid advancements in computer vision, ship detection models based on deep learning have been more and more prevalent. However, most network methods use expensive costs with high hardware equipment needed to increase detection accuracy. In response to this challenge, a lightweight real-time detection approach called YOLOSeaShip is proposed. Firstly, derived from the YOLOv7-tiny model, the partial convolution was utilized to replace the original 3×1 convolution in the ELAN module to further fewer parameters and improve the operation speed. Secondly, the parameter-free average attention module was integrated to improve the locating capacity for the hull of a ship in an image. Finally, the accuracy changes of the Focal EIoU hybrid loss function under different parameter changes were studied. The practical results trained on the SeaShips (7000) dataset demonstrate that the suggested method can detect and classify the ship position from the image more efficiently, with mAP of 0.976 and FPS of 119.84, which is ideal for real-time ship detection applications.

Transfer learning to leverage larger datasets for improved prediction of protein stability changes

Amino acid mutations that lower a protein's thermodynamic stability are implicated in numerous diseases, and engineered proteins with enhanced stability can be important in research and medicine. Computational methods for predicting how mutations perturb protein stability are, therefore, of great interest. Despite recent advancements in protein design using deep learning, in silico prediction of stability changes has remained challenging, in part due to a lack of large, high-quality training datasets for model development. Here, we describe ThermoMPNN, a deep neural network trained to predict stability changes for protein point mutations given an initial structure. In doing so, we demonstrate the utility of a recently released megascale stability dataset for training a robust stability model. We also employ transfer learning to leverage a second, larger dataset by using learned features extracted from ProteinMPNN, a deep neural network trained to predict a protein's amino acid sequence given its three-dimensional structure. We show that our method achieves state-of-the-art performance on established benchmark datasets using a lightweight model architecture that allows for rapid, scalable predictions. Finally, we make ThermoMPNN readily available as a tool for stability prediction and design.

MixMobileNet: A Mixed Mobile Network for Edge Vision Applications

Currently, vision transformers (ViTs) have rivaled comparable performance to convolutional neural networks (CNNs). However, the computational demands of the transformers’ self-attention mechanism pose challenges for their application on edge devices. Therefore, in this study, we propose a lightweight transformer-based network model called MixMobileNet. Similar to the ResNet block, this model only comprises a MixMobile block (MMb), which combines the efficient local inductive bias with the explicit modeling features of a transformer to achieve the fusion of the local–global feature interactions. For local, we propose the local-feature aggregation encoder (LFAE), which incorporates a PC2P (Partial-Conv→PWconv→PWconv) inverted bottleneck structure for residual connectivity. In particular, the kernel and channel scale are adaptive, reducing feature redundancy in adjacent layers and efficiently representing parameters. For global, we propose the global-feature aggregation encoder (GFAE), which employs a pooling strategy and computes the covariance matrix between channels instead of the spatial dimensions, changing the computational complexity from quadratic to linear, and this accelerates the inference of the model. We perform extensive image classification, object detection, and segmentation experiments to validate model performance. Our MixMobileNet-XXS/XS/S achieves 70.6%/75.1%/78.8% top-1 accuracy with 1.5 M/3.2 M/7.3 M parameters and 0.2 G/0.5 G/1.2 G FLOPs on ImageNet-1K, outperforming MobileViT-XXS/XS/S with an improvement of +1.6%↑/+0.4%↑/+0.4%↑ with −38.8%↓/−51.5%↓/−39.8%↓ reduction in FLOPs. In addition, the MixMobileNet-S assembly of SSDLite and DeepLabv3 achieves an accuracy of 28.5 mAP/79.5 mIoU at COCO2017/VOC2012 with lower computation, demonstrating the competitive performance of our lightweight model.

TinyML: Adopting tiny machine learning in smart cities

<div><p>Since Tiny machine learning (TinyML) is a quickly evolving subject, it is crucial that internet of things (IoT) devices be able to communicate with one another for the sake of stability and future development. TinyML is a rapidly growing subfield at the intersection of computer science, software engineering, and machine learning. Building deep learning (DL) networks that are a few hundred KBs in size has been the focus of recent research in this area. Deploying TinyML into small devices makes them smart. Reduced computation, power usage, and response time are just a few of the many advantages of TinyML. In this work, we provide an introduction to TinyML and demonstrate its benefits and architecture. Then, we investigate the meaning of quantization as a standard compression method for TinyML-related applications. There are two methods used to obtain the quantized weights of the deep learning models are quantization-aware training (QAT) and post-training quantization (PTQ), we described them in details. Next, TinyML-based solutions to improve the role of IoT devices in Smart Cities are highlighted as: lightweight training of deep learning models, inference of lightweight deep learning models in IoT devices, low power consumption, and inference of deep learning models in restricted resources of IoT devices. Finally, presenting some use cases for TinyML studies, with these studies applied to several cases in a variety of fields. To the best of the author’s knowledge, few studies have investigated TinyML as it is an emerging field.</p></div>