Lightweight Convolutional Neural Network Research Articles

Artificial Intelligence-Driven Detection of LASIK Using Corneal Optical Coherence Tomography Maps.

To train and validate a convolutional neural network (CNN) to detect the history of laser-assisted in situ keratomileusis (LASIK) surgeries using corneal optical coherence tomography (OCT) maps. Five corneal OCT maps (pachymetry, epithelial thickness, posterior mean curvature, anterior axial power, and anterior stroma reflectance) were utilized as the input of a lightweight CNN model. OCT scans of healthy volunteers and patients who had undergone myopic or hyperopic LASIK were included. Repeated fivefold cross-validation was used to train and evaluate the proposed CNN. In addition, a separate group of post-LASIK participants, who were not included in the cross-validation, was used for out-of-sample testing to assess the CNN model performance. In the cross-validation, the proposed CNN model achieved an overall balanced accuracy of 90.2% ± 3.6% with 93.5% ± 5.2% sensitivity and 97.8% ± 1.7% area under the receiver operating characteristic curve (AUC) in detecting myopic LASIK and 90.2% ± 5.8% sensitivity and 98.2% ± 1.9% AUC in identifying the hyperopic LASIK. In the out-of-sample test, all eyes were classified correctively. The lightweight CNN model with corneal OCT maps provides a useful tool for detecting LASIK history. Artificial intelligence-assisted OCT may offer better management for patients with LASIK history who need cataract surgeries.

A lightweight convolutional neural network for surface defect detection in strip steel

Abstract In the context of resource constrained embedded terminals, the current deep learning-based technology for detecting surface defects in strip steel faces challenges with low detection accuracy and efficiency. This study introduces a lightweight convolution neural network (CNN) model for detecting surface defects of strip steel, which is based on deep learning. Through the integration of non-parameter attention mechanism SimAM, inverse residual structure, and depth-wise convolution module, a lightweight attention mobile (LAM) backbone network is developed to achieve optimal feature extraction. To address the challenge of detecting small surface defects the context augmentation module (CAM) is introduced to provide more information for small defects detection by using multi-scale features. To improve the efficiency of feature fusion and reduce parameter redundancy, the GVFPN neck network is proposed. The network aims to represent and deal with multi-scale features objectively while minimizing costs. The experimental results indicate that on the NEU-DET dataset, the proposed network attains an mAP of 78.8\% and FPS at 97 frame/s. \textcolor{red}{Moreover, the model requires only 2.39 M parameters and 3.1 G FLOPs. Compared to YOLOv5s, the proposed network reduces parameters by 65.9\% and FLOPs by 80.7\%, while achieving a 1.3\% higher mAP and a 30 frame/s increase in detection speed. These results effectively demonstrate the effectiveness and superiority of the proposed method.

A Lightweight Network Based on Dynamic Split Pointwise Convolution Strategy for Hyperspectral Remote Sensing Images Classification

For reducing the parameters and computational complexity of networks while improving the classification accuracy of hyperspectral remote sensing images (HRSIs), a dynamic split pointwise convolution (DSPC) strategy is presented, and a lightweight convolutional neural network (CNN), i.e., CSM-DSPCss-Ghost, is proposed based on DSPC. A channel switching module (CSM) and a dynamic split pointwise convolution Ghost (DSPC-Ghost) module are presented by combining the presented DSPC with channel shuffling and the Ghost strategy, respectively. CSM replaces the first expansion pointwise convolution in the MobileNetV2 bottleneck module to reduce the parameter number and relieve the increasing channel correlation caused by the original channel expansion pointwise convolution. DSPC-Ghost replaces the second pointwise convolution in the MobileNetV2 bottleneck module, which can further reduce the number of parameters based on DSPC and extract the depth spectral and spatial features of HRSIs successively. Finally, the CSM-DSPCss-Ghost bottleneck module is presented by introducing a squeeze excitation module and a spatial attention module after the CSM and the depthwise convolution, respectively. The presented CSM-DSPCss-Ghost network consists of seven successive CSM-DSPCss-Ghost bottleneck modules. Experiments on four measured HRSIs show that, compared with 2D CNN, 3D CNN, MobileNetV2, ShuffleNet, GhostNet, and Xception, CSM-DSPCss-Ghost can significantly improve classification accuracy and running speed while reducing the number of parameters.

MSNet: Food Image Classification Model Based on Lightweight Convolutional Neural Network

With the advancement of technology, the demand for healthy eating has increased, making food classification a research hotspot. Existing deep learning-based food image classification models demonstrate high accuracy but require substantial computational resources, limiting their use on resource-constrained devices. In this study, a lightweight convolutional neural network model named MSNet for food classification is proposed. MSNet mainly consists of M Blocks and S Blocks. The M Block uses improved depthwise convolution to reduce the computational cost of conventional convolutions, and the S Block uses channel shuffle techniques to enhance feature information flow between channels without increasing additional computation, effectively capturing relationships between different channel features. Experimental results on three benchmark datasets (ETHZ Food-101, Vireo Food-172, and ISIA Food-500) show that MSNet achieves top-1 accuracies of 86.24%, 87.98%, and 65.70%, with model sizes of 13.8 MB, 15.9 MB, and 25.4 MB, respectively, outperforming mainstream models in terms of computational efficiency. Further quantization produces two MSNet-Lite variants with competitive model size, while maintaining high accuracy and significantly improving inference speed. Additionally, visualization analysis indicates that MSNet effectively extracts essential features of food images, offering good interpretability and generalization across datasets of varying complexity. The proposed MSNet model provides a feasible solution for practical deployment in food classification tasks on mobile and embedded devices.

Research on Crop Disease Image Recognition Based on Internet of Things Technology and Stacking Integrated Learning

ABSTRACTIn the field of agriculture, disease control and management have been a hot research topic of great interest. In recent years, with the reduction of the cost of image sensors and the improvement of the accuracy of deep‐learning algorithms, various information processing methods have been widely used in agricultural production. In this paper, an in‐depth exploration of crop disease image recognition methods based on IoT technology is carried out. Initially, an innovative method of deploying sensor nodes within an irregular triangular grid is designed to facilitate effective data collection. Subsequently, accurate image segmentation and feature extraction were executed on the accumulated data. A two‐tier Stacking framework was used to integrate three lightweight convolutional neural networks. The first level classifier is used to generate data output values for model training; the second level classifier learns further from the output of the first level classifier, corrects the bias of each individual learner in the framework, and produces the final prediction. On the publicly available PlantVillage data set, the EMNet integration model proposed in this thesis has an accuracy of 98.96%, which is at least 0.68% better than other influential DCNN validation accuracies, with good robustness and generalization.

HSCFNet: Lightweight Convolutional Neural Network for the Classification of Infectious and Non‐Infectious Skin Diseases

<scp>HSCFNet</scp>: Lightweight Convolutional Neural Network for the Classification of Infectious and Non‐Infectious Skin Diseases

Tomato leaf disease detection using Taguchi-based Pareto optimized lightweight CNN

&lt;p&gt;The prospect of food security becoming a global danger by 2050 due to the exponential growth of the world population. An increase in production is indispensable to satisfy the escalating demand for food. Considering the scarcity of arable land, safeguarding crops against disease is the best alternative to maximize agricultural output. The conventional method of visually detecting agricultural diseases by skilled farmers is time-consuming and vulnerable to inaccuracies. Technology-driven agriculture is an integral strategy for effectively addressing this matter. However, orthodox lightweight convolutional neural network (CNN) models for early crop disease detection require fine-tuning to enhance the precision and robustness of the models. Discovering the optimal combination of several hyperparameters might be an exhaustive process. Most researchers use trial and error to set hyperparameters in deep learning (DL) networks. This study introduces a new systematic approach for developing a less sensitive CNN for crop leaf disease detection by hyperparameter tuning in DL networks. Hyperparameter tuning using a Taguchi-based orthogonal array (OA) emphasizes the S/N ratio as a performance metric primarily dependent on the model’s accuracy. The multi-objective Pareto optimization technique accomplished the selection of a robust model. The experimental results demonstrated that the suggested approach achieved a high level of accuracy of 99.846% for tomato leaf disease detection. This approach can generate a set of optimal CNN models’ configurations to classify leaf disease with limited resources accurately.&lt;/p&gt;

A Lightweight CNN Model for Vision Based Fire Detection on Embedded Systems

Existing fire detection systems are generally sensor-based. Such traditional sensor-based systems have limited detection range, slow response, high false alarm and inability to give rich descriptive information. To overcome these shortcomings, computer vision-based methods for fire detection have been proposed. These vision-based systems have advantages of faster response, larger surveillance coverage, affordable cost and less human interference. However, the performance of such systems is affected by the complexity of scene under observation, irregular lighting and low-quality frames. The limitations of both traditional sensor-based and vision-based systems for fire detection can be addressed by using convolutional neural network (CNN). Despite their superior performance in various computer vision tasks, computational complexity of CNN models remain a key concern for deployment on embedded platforms which are characterised by limited resources and memory. This paper presents a computationally less expensive CNN model for fire detection that can be easily deployed on embedded hardware platforms such as FPGA. SqueezeNet, a pre-trained CNN model on ImageNet dataset, was modified and trained using transfer learning approach to classify fire images. SqueezeNet is a computationally less intensive CNN architecture which is 18 layer deep and 5.2 MB in size that makes it a good choice for embedded applications. It yielded an accuracy of 95% on the benchmark dataset which is better than state-of-the-art feature-based approaches. But the performance of the model can still be enhanced through further experimentation by considering a larger dataset.

A Robust Visual-Inertial Navigation Method for Illumination-Challenging Scenes

Visual-inertial odometry (VIO) has been found to have great value in robot positioning and navigation. However, the existing VIO algorithms rely heavily on excellent lighting environments and the accuracy of robot positioning and navigation is degraded largely in illumination-challenging scenes. A robust visual-inertial navigation method is developed in this paper. We construct an effective low-light image enhancement model using a deep curve estimation network (DCE) and a lightweight convolutional neural network to recover the texture information of dark images. Meanwhile, a brightness consistency inference method based on the Kalman filter is proposed to cope with illumination variations in image sequences. Multiple sequences obtained from UrbanNav and M2DRG datasets are used to test the proposed algorithm. Furthermore, we also conduct a real-world experiment for the proposed algorithm. Both experimental results demonstrate that our algorithm outperforms other state-of-art algorithms. Compared to the baseline algorithm VINS-mono, the tracking time is improved from 22.0% to 68.2% and the localization accuracy is improved from 0.489[Formula: see text]m to 0.258[Formula: see text]m on the darkest sequences.

Hybrid deep learning model for predicting failure properties of asphalt binder from fracture surface images

AbstractCracking impacts asphalt concrete durability primarily due to cohesive asphalt binder failures. The poker chip test has recently been introduced to better characterize the cracking potential of asphalt binders by fracturing a specimen in a realistic stress state to a thin binder film. However, broader adoption faces challenges due to high instrumentation costs for measuring load and displacement. This paper presents and validates a deep learning model that predicts ductility and tensile strength from posttest images of fractured binder surfaces, with potential extensions to simplified instrumentation. The hybrid model, named PCNet, integrates a custom lightweight convolutional neural network (CNN) developed to capture local features (e.g., edges, boundaries, contours) within fracture cavities with a Swin Transformer that models global contextual dependencies. A bidirectional cross‐attention fusion module is designed to facilitate mutual information exchange between CNN and transformer branches. The fused features are then processed by a fully connected network (FCN) to predict indices derived from the test. The proposed model demonstrates high predictive accuracy across a range of binders and test configurations, achieving an of 0.966 and a mean absolute percentage error (MAPE) of 12.95% in predicting ductility, while also attaining an of 0.947 and a MAPE of 9.15% for strength, outperforming standalone models. Monte Carlo Dropout is also incorporated in the FCN to quantify prediction confidence. This cost‐effective methodology provides insights into fracture propagation in soft viscoelastic media and contributes to the field of experimental mechanics. With further data collection, the model holds potential for broader implementation, directly linking fracture surface images and mixture or field‐scale cracking behavior.

Comprehensive Serum Analysis via an AI-Assisted Magnetically Driven SERS Platform for the Diagnosis and Etiological Differentiation of Childhood Epilepsy.

Timely and accurate diagnosis of childhood epilepsy and identification of its etiology are crucial for early interventional treatment, yet still, effective detection methods are lacking. Blood analysis is a promising strategy for disease diagnosis. However, due to the complex composition and lack of definite childhood epilepsy diagnostic markers in serum, comprehensively profiling serum molecular signals to accurately reveal diagnostic information is still challenging. Herein, we developed a novel magnetically driven SERS platform, which utilized specially designed branched Au nanostructure-embedded magnetic microspheres to achieve simultaneous detection of small molecules and biomacromolecules in serum, thus providing comprehensive serum molecular SERS signals. By using this platform, the SERS data sets of serum samples from 90 healthy controls and 585 epileptic patients were collected to train a self-built lightweight convolutional neural network (MLS-CNN) model, which successfully identified the serum epileptic diagnostic and etiological differentiation information, including causes of autoimmune encephalitis, febrile infection, developmental disability, structural brain lesions, and unknown etiology. The MLS-CNN model exhibits excellent diagnostic accuracy (100%) and etiological differentiation accuracy (>89%) for epilepsy. This AI-assisted magnetically driven SERS platform for comprehensively profiling the molecular information on serum might provide a novel strategy for childhood epilepsy diagnosis and etiological identification.

Real-Time Damage Detection in Fiber Lifting Ropes Using Lightweight Convolutional Neural Networks

Real-Time Damage Detection in Fiber Lifting Ropes Using Lightweight Convolutional Neural Networks

Developing a Lightweight CNN-Based Model for Enhanced Automatic White Blood Cell Classification Using Deep Learning Techniques

Finding and monitoring many diseases, including infections and leukaemia, depend on the classification of white blood cells (WBCs). Conventional techniques mostly rely on seeing samples under a microscope and calling for an expert to evaluate the outcomes. This method takes a lot of time and human mistake might create errors. This work uses a lightweight convolutional neural network (CNN) to provide a basic model to enhance the automated categorisation of white blood cells (WBCs) via deep learning. The proposed paradigm maximises computational efficiency, hence it is perfect for real-time applications and environments with limited resources. We kept the model correct while still designing a simplified form that would make using it easier. Trained and validated on a whole dataset including diverse microscopy images of WBCs, the model guarantees stability and generalizability throughout many imaging environments. Our findings reveal that, in terms of sorting accuracy, speed of processing, and operational efficiency, we have achieved significant progress over conventional techniques. The minimal weight of the device makes it simple to use in internet services and mobile health applications, therefore enabling more individuals to get necessary medical attention. By means of concepts on building efficient deep learning models for healthcare, this work provides a beneficial tool for better categorising white blood cells and also helps enhance medical photographs.

LiteDiagNet: Enhancing noise-resistant fault diagnosis in rotating machinery via IoT-Based cloud-edge collaboration

Rapid advances in the intelligent Internet of Things (IoT) have significantly enhanced sensor technology and edge computing across various monitoring systems. This study focuses on the seamless integration of intelligent IoT with vibration sensor technology and edge computing to effectively detect faults in rotating machinery amidst noisy environments. We introduce an innovative framework for fault diagnosis through cloud-edge collaboration, aimed at broadening the applications of intelligent IoT. Central to our approach is LiteDiagNet, a novel, lightweight Convolutional Neural Network (CNN) tailored to identify bearing faults in noise-affected signals. LiteDiagNet features a cutting-edge architecture with base blocks and a distinctive “LiteBlock” that employs multiscale, depth-wise convolutions to significantly improve performance. We evaluated our model’s efficacy against several lightweight networks using public datasets and verified its robustness in real-world scenarios by simulating two types of noise at five different intensity levels. The results demonstrate that LiteDiagNet achieves superior accuracy with fewer parameters and reduced computational demands, thereby enhancing efficiency. This research not only offers a groundbreaking solution for fault diagnosis in rotating machinery but also advances the intelligent and efficient application of industrial IoT technologies.

Lightweight CNN-Based Visual Perception Method for Assessing Local Environment Complexity of Unmanned Surface Vehicle.

Addressing the problem of inadequate environmental detection in the process of optimizing search for unmanned surface vehicles (USVs) by a heuristic algorithm, this paper proposes a comprehensive visual perception method that combines a lightweight convolutional neural network (CNN) with the USV's real-time heading angle. This method employs a multi-feature input CNN with residual learning blocks, which takes both the current local environmental images and heading angle features as inputs to identify the complexity of the local environment with higher accuracy and a smaller load size. Meanwhile, human expertise is incorporated to classify labels through a majority voting system, thereby making the model's perceptual classification more intuitive and allowing it to possess a human-like comprehensive perception ability compared to systems with classification methods with several parameters. Subsequently, this identification result can be used as feedback for the heuristic algorithm to optimize and plan the USV's path. The simulation results indicate that the developed model achieves an 80% reduction in model size while maintaining an accuracy exceeding 90%. The proposed method significantly improves the environment recognition capability of the heuristic algorithm, enhances optimization search efficiency, and increases the overall performance of path planning by approximately 21%.

Efficient Design and Compression of CNN Models for Rapid Character Recognition

Convolutional Neural Networks (CNNs) are extensively utilized for image processing and recognition tasks; however, they often encounter challenges related to large model sizes and prolonged training times. These limitations present difficulties in resource-constrained environments that require rapid model deployment and efficient computation. This study introduces a systematic approach to designing lightweight CNN models specifically for character recognition, emphasizing the reduction of model complexity, training duration, and computational costs without sacrificing performance. Techniques such as hyperparameter tuning, model pruning, and post-training quantization (PTQ) are employed to decrease model size and enhance training speed. The proposed methods are particularly well-suited for deployment on edge computing platforms, such as Raspberry Pi, or embedded systems with limited resources. Our results demonstrate a reduction of over 80% in model size, decreasing from 43.73 KB to 6.25 KB, and a reduction of more than 45% in training time, decreasing from over 150 seconds to less than 80 seconds. This research highlights the potential for achieving a balance between efficiency and accuracy in CNN design for real-world deployment, addressing the increasing demand for streamlined deep learning models in resource-constrained environments.

An intelligent fruit freshness monitoring system using hydrophobic indicator labels based on methylcellulose, k-carrageenan, and sodium tripolyphosphate, combined with deep learning.

An intelligent fruit freshness monitoring system using hydrophobic indicator labels based on methylcellulose, k-carrageenan, and sodium tripolyphosphate, combined with deep learning.

A lightweight CNN model for UAV-based image classification

For many unmanned aerial vehicle (UAV)-based applications, especially those that need to operate with resource-limited edge networked devices in real-time, it is crucial to have a lightweight computing model for data processing and analysis. In this study, we focus on UAV-based forest fire imagery detection using a lightweight convolution neural network (CNN). The task is challenging owing to complex image backgrounds and insufficient training samples. Specifically, we enhance the MobileNetV2 model with an attention mechanism for UAV-based image classification. The proposed model first employs a transfer learning strategy that leverages the pre-trained weights from ImageNet to expedite learning. Then, the model incorporates randomly initialised weights and dropout mechanisms to mitigate over-fitting during training. In addition, an ensemble framework with a majority voting scheme is adopted to improve the classification performance. A case study on forest fire scenes classification with benchmark and real-world images is demonstrated. The results on a publicly available UAV-based image data set reveal the competitiveness of our proposed model as compared with those from existing methods. In addition, based on a set of self-collected images with complex backgrounds, the proposed model illustrates its generalisation capability to undertake forest fire classification tasks with aerial images.

Defect detection in fused deposition modelling using lightweight convolutional neural networks

Defect detection in fused deposition modelling using lightweight convolutional neural networks

Lung and Colon Cancer Classification Using Multiscale Deep Features Integration of Compact Convolutional Neural Networks and Feature Selection

The automated and precise classification of lung and colon cancer from histopathological photos continues to pose a significant challenge in medical diagnosis, as current computer-aided diagnosis (CAD) systems are frequently constrained by their dependence on singular deep learning architectures, elevated computational complexity, and their ineffectiveness in utilising multiscale features. To this end, the present research introduces a CAD system that integrates several lightweight convolutional neural networks (CNNs) with dual-layer feature extraction and feature selection to overcome the aforementioned constraints. Initially, it extracts deep attributes from two separate layers (pooling and fully connected) of three pre-trained CNNs (MobileNet, ResNet-18, and EfficientNetB0). Second, the system uses the benefits of canonical correlation analysis for dimensionality reduction in pooling layer attributes to reduce complexity. In addition, it integrates the dual-layer features to encapsulate both high- and low-level representations. Finally, to benefit from multiple deep network architectures while reducing classification complexity, the proposed CAD merges dual deep layer variables of the three CNNs and then applies the analysis of variance (ANOVA) and Chi-Squared for the selection of the most discriminative features from the integrated CNN architectures. The CAD is assessed on the LC25000 dataset leveraging eight distinct classifiers, encompassing various Support Vector Machine (SVM) variants, Decision Trees, Linear Discriminant Analysis, and k-nearest neighbours. The experimental results exhibited outstanding performance, attaining 99.8% classification accuracy with cubic SVM classifiers employing merely 50 ANOVA-selected features, exceeding the performance of individual CNNs while markedly diminishing computational complexity. The framework’s capacity to sustain exceptional accuracy with a limited feature set renders it especially advantageous for clinical applications where diagnostic precision and efficiency are critical. These findings confirm the efficacy of the multi-CNN, multi-layer methodology in enhancing cancer classification precision while mitigating the computational constraints of current systems.