Lightweight Deep Neural Network Research Articles

Autonomous path planning with obstacle avoidance for smart assistive systems

Given the increased interest in smart assistive technologies and autonomous robot vehicles, path planning has emerged as one of the most researched and challenging topics in navigation. Moving to partially known or unknown environment, an assistive navigation system should be able to extract spatiotemporal information and dynamically identify objects and adjust the route. Current approaches typically rely on external services to perform high demanding computations and employ a plethora of overlapping sensors to accurately scan the surrounding environment. This increases their energy demands, size and weight, while incommodes their use in real time applications making their application to wearable assistive systems, such as smart glasses, a challenge. Aiming to provide a comfortable and computationally efficient wearable solution that can be used by human or robotic assistive systems, in this study we propose a novel two-level hierarchical architecture combining global and local path planning. The macroscale navigation involves the construction of the initial global path while the microscale navigation includes the local path planning with obstacle detection and avoidance. The methodology consists of: (i) a novel chaotic ant colony optimization algorithm with fuzzy logic (CACOF) for path construction; (ii) powerful and light weight deep convolutional neural networks for obstacle detection; and (iii) a Bug-like algorithm enhanced with fuzzy rules for obstacle avoidance in case of static objects. A vast experimental evaluation was conducted to test the proposed methodologies in a simulation environment based on the topology of real area. The results proved the computational efficiency and ability of the proposed path planning algorithms to address effectively multi-objective global and path planning problems which make them suitable for real time applications.

Watch out, pothole! featuring road damage detection in an end-to-end system for autonomous driving

While autonomous driving technology made significant progress in the last decade, road damage detection as a relevant challenge for ensuring safety and comfort is still under development. This paper addresses the lack of algorithms for detecting road damages that meet autonomous driving systems’ requirements. We investigate the environmental perception systems’ architecture and current algorithm designs for road damage detection. Based on the autonomous driving architecture, we develop an end-to-end concept that leverages data from low-cost pre-installed sensors for real-time road damage and damage severity detection as well as cloud- and crowd-based HD Feature Maps to share information across vehicles. In a design science research approach, we develop three artifacts in three iterations of expert workshops and design cycles: the end-to-end concept featuring road damages in the system architecture and two lightweight deep neural networks, one for detecting road damages and another for detecting their severity as the central components of the system. The research design draws on new self-labeled automotive-grade images from front-facing cameras in the vehicle and interdisciplinary literature regarding autonomous driving architecture and the design of deep neural networks. The road damage detection algorithm delivers cutting-edge performance while being lightweight compared to the winners of the IEEE Global Road Damage Detection Challenge 2020, which makes it applicable in autonomous vehicles. The road damage severity algorithm is a promising approach, delivering superior results compared to a baseline model. The end-to-end concept is developed and evaluated with experts of the autonomous driving application domain.

Knowledge distillation technique enabled hardware efficient OSNR monitoring from directly detected PDM-QAM signals

Deep neural network (DNN) enabled optical SNR (OSNR) monitoring is vital to facilitate the management of agile optical networks. However, both the scale size and the implementation complexity of DNNs prevent their practical application. Here, we experimentally demonstrate the knowledge distillation technique (KDT) for DNN-based OSNR monitoring from the directly detected polarization division multiplexed 16 quadrature amplitude modulation (PDM-16QAM) and PDM-64QAM signals by the use of amplitude histograms. First, both floating-point operations (FLOPs) and parameters (Params) are introduced into the fiber optic communications as the evaluation metrics for the computational complexity of DNNs. Then, the pre-trained teacher network and the to-be-used student network constitute a KDT network, while the task of the pre-trained teacher network transfers the knowledge to the student network. From the experimental results, we learn that the KDT benefits the student network and realizes a rms error (RMSE) of 0.98 dB over OSNR ranges of 14–24 dB and 23–34 dB for 10 Gbaud PDM-16QAM and PDM-64QAM signals, respectively. The Params of the student network are only 0.004M, and the FLOPs are 0.25M, which are less than that of existing OSNR monitoring schemes. Meanwhile, an RMSE improvement of 0.23 dB arising in the OSNR monitoring has been secured to the student network with the help of KDT. Therefore, we believe our proposed KDT is promising for optical performance monitoring with a lightweight DNN.

An Enhanced Competitive Swarm Optimizer With Strongly Convex Sparse Operator for Large-Scale Multiobjective Optimization

Sparse multiobjective optimization problems (MOPs) have become increasingly important in many applications in recent years, e.g., the search for lightweight deep neural networks and high-dimensional feature selection. However, little attention has been paid to sparse large-scale MOPs, whose Pareto-optimal sets are sparse, i.e., with many decision variables equal to zero. To address this issue, this article proposes an enhanced competitive swarm optimization algorithm assisted by a strongly convex sparse operator (SCSparse). A tricompetition mechanism is introduced into competitive swarm optimization, aiming to strike a better balance between exploration and exploitation. In addition, the SCSparse is embedded in the position updating of the particles to generate sparse solutions. Our simulation results show that the proposed algorithm outperforms the state-of-the-art methods on both sparse test problems and application examples.

Specializing CGRAs for Light-Weight Convolutional Neural Networks

Deep neural network (DNN) processing units, or DPUs, are one of the most energy-efficient platforms for DNN applications. However, designing new DPUs for every DNN model is very costly and time consuming. In this article, we propose an alternative approach: to specialize coarse-grained reconfigurable architectures (CGRAs), which are already quite capable of delivering high performance and high energy efficiency for compute-intensive kernels. We identify a small set of architectural features on a baseline CGRA to enable high-performance mapping of depthwise convolution (DWC) and pointwise convolution (PWC) kernels, which are the most important building block in recent light-weight DNN models. Our experimental results using MobileNets demonstrate that our proposed CGRA enhancement can deliver <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$8\sim 18\times $ </tex-math></inline-formula> improvement in area-delay product (ADP) depending on layer type, over a baseline CGRA with a state-of-the-art CGRA compiler. Moreover, our proposed CGRA architecture can also speed up 3-D convolution with similar efficiency as previous work, demonstrating the effectiveness of our architectural features beyond depthwise separable convolution (DSC) layers.

DNN-CBAM: An enhanced DNN model for facial emotion recognition

Facial expression recognition is a current research hotspot and can be applied to computer vision fields such as human-computer interaction and affective computing. The lack of diversity and category recognition information in the neural network input may affect the performance of the network, resulting in insufficient extraction of facial expression features. In order to address the above problems, a lightweight deep convolution neural network with convolution block attention module is proposed in this paper. The implementation of the lightweight DNN relies on the use of deep separable convolution and residual blocks. The combination of the convolution block attention module and the improved classification function can optimize the lightweight model. We use accuracy and confusion matrix to evaluate different models, ultimately achieving 71.5% and 99.5% accuracy on the Fer2013 and CK+ datasets respectively. The experimental results show that our model has good feature representation capabilities.

Bare-hand gesture occlusion-aware interactive augmented reality assembly

Augmented reality (AR) has increasingly been applied to benefit manual assembly tasks by conveying the paper-based instructions into visual guidance intuitively, alleviating the cognitive load while improving assembly efficiency. Nevertheless, most current AR assemblies focus on spatial alignment by superimposing a 3D virtual model on the real scene directly, ignoring the occlusion awareness between virtual guidance and the real scene. To this end, we proposed a bare-hand occlusion-aware interactive AR assembly method based on the monocular image, and a lightweight deep neural network is established to infer the depth relationship between the 3D virtual model and real scene including gesture manipulation, and thus the ambiguous AR instruction by inaccurate occlusion deduction would be prevented, leading to more realistic bare-hand interactive AR guidance for manual assembly. And then, a quantitative evaluation criterion is established to illustrate the gesture occlusion awareness performance during manual assembly operations. Finally, comprehensive experiments are carried out and the results illustrate that the bare-hand occlusion-aware system can alleviate the cognitive load for the operators within interactive AR assembly tasks, providing a more human-center intelligent assembly application.

IMCLNet: A lightweight deep neural network for Image-based Malware Classification

With the increasing number of malware and advanced evasion technology, it is more and more difficult to detect malware accurately and efficiently. To solve this challenge, a feasible method is to convert malware into images, and then classify them by using the model based on a convolutional neural network. However, due to the highly imbalanced datasets, image-based methods generally rely on data enhancement or pre-training parameters, which makes the classification process not lightweight enough. Meanwhile, most of these methods lack a detailed study of the image size during the process of conversion. To achieve an accurate and efficient classification, we propose a lightweight malware classification model, IMCLNet, which is driven by malware images and does not need feature engineering and domain knowledge. When designing the model, we comprehensively weighed accuracy, the calculation cost, and the number of parameters, and integrated Coordinate Attention, Depthwise Separable Convolution, and Global Context Embedding. We evaluated IMCLNet on two large datasets, MalImg and BIG2015, and without data enhancement and pre-training parameters, our proposed method still achieved 99.785% and 98.942% classification accuracy. IMCLNet predicts that a malware image of size 32 × 32 takes only 0.95ms and 0.84ms, respectively. In addition, we also compare IMCLNet with the mainstream lightweight models such as MobileNetV3, ShuffleNetV2, and MixNet. The experimental results show that IMCLNet has obvious advantages in training time, accuracy, the number of parameters, model size, and prediction time on GPU.

DGAFF: Deep genetic algorithm fitness Formation for EEG Bio-Signal channel selection

Brain-computer interface systems aim to facilitate human–computer interactions by directly translating brain signals for computers. Recently, using many electrodes has led to better performance in these systems. However, increasing the number of recorded electrodes causes additional time, hardware, and computational costs besides undesired complications of the recording process. Channel selection decreases data dimension and eliminates irrelevant channels while reducing the noise effects. Furthermore, the technique lowers the time and computational costs in the test phase. We present a channel selection method, which combines a sequential search method with a genetic algorithm called Deep Genetic Algorithm Fitness Formation (DGAFF). The proposed method accelerates the convergence of the genetic algorithm and increases the system’s performance. The system evaluation is based on a lightweight deep neural network that automates the whole model training process. Our method, compared to other channel selection methods, outperforms the tradeoff between the classification accuracy and the number of selected channels.

Cooperative Inference of DNNs for Delay- and Memory-Constrained Wireless IoT Systems

This work studies the cooperative inference of deep neural networks (DNNs), in which a memory-constrained end device performs a delay-constrained inference process with an aid of an edge server. Although several works considered the cooperative inference of DNNs in the literature, it was assumed in those works that the memory footprints at end devices are unlimited, which is in practice not realistic. To address this issue, in this work, a memory-aware cooperative DNN inference is proposed. Specifically, we propose to adopt knowledge distillation to obtain high-performing lightweight DNNs. To minimize the inference delay, we first analyze the end-to-end delay required for processing the proposed cooperative DNN inference, and then we minimize the delay by jointly optimizing the DNN partitioning point and the intermediate data transmission rate. Also, a dynamic DNN selection scheme is developed by fully exploiting the available memory resource in order to maximize the performance of the inference task in terms of inference accuracy. Experimental results demonstrate that the proposed cooperative DNN inference considerably outperforms the comparable schemes while satisfying both the delay constraint and the memory constraint.

Internal defects detection and classification in hollow cylindrical surfaces using single shot detection and MobileNet

Surface defect inspection, detection, and classification in hollow cylindrical surfaces such as pipes and barrels have a significant impact on the structural integrity of various industrial products. Regular inspection and identification of the faults reduces the likelihood of faults’ aggravation, limits the damaging effects, and increases the product life. However, most of the defect detection algorithms for cylindrical surfaces rely heavily on handcrafted feature extraction limiting the ability to recognize the defects effectively. This research work proposes an image processing-based automatic defect detection and classification approach for cylindrical hollow surfaces. The proposed system uses a single shot multi-box detection (SSD) algorithm for localization and a customized lightweight deep convolutional neural network as a backbone network to classify defects generally found in industrial pipes and gun barrels. First, the image dataset is acquired from a real-time working environment using an indigenously developed borescope featuring a rotating camera and special hardware features. Later, the bounding boxes are calculated using extracted features to localize defects with SSD which takes a single shot to detect multiple objects within the image. Finally, the defected regions are classified into five classes of commonly found issues of pitting, chipping, rusting, dirt, and thermal cracking by utilizing deep learning architecture of 53 layers. It is found that the proposed approach can indicate the exact location of the classified defect in terms of angle and distance from a reference point. Also, the proposed method improves the detection and classification accuracy significantly compared to other existing methods. To encourage the development and evaluation of new methods for cylindrical surface defect detection, the dataset is also made publicly available.

An Efficient License Plate Detection Approach Using Lightweight Deep Convolutional Neural Networks

Benefited from deep convolutional neural networks, various license plate detection methods based on deep networks have been proposed and achieved significant improvements compared with traditional methods. However, the high computational cost due to complex structures prevents these methods from being deployed in real-world applications. This paper proposes an efficient license plate detection method based on lightweight deep convolutional neural networks for improving the detection speed. To extract high-level features from input images, this paper designs a lightweight feature pyramid generation module based on a lightweight architecture and depth-wise convolutions. To further enhance feature pyramid, an efficient feature enhancement module is designed to fuse features generated by the region proposal network with backbone features. In the detection network, a light head structure based on fully connected layers is employed to further reduce the computational cost of the model. In experiments, floating point operations and detection ratio are used to evaluate the efficient of the proposed method. Experimental results on public datasets show that the proposed method achieves the best trade-off between speed and accuracy.

A lightweight deep neural network implemented on MATLAB without using GPU for the automatic monitoring of the plants

A lightweight deep neural network implemented on MATLAB without using GPU for the automatic monitoring of the plants

Image classification method based on local receptive field extended S-Mobilenet model

Aiming at the problem that the lightweight deep neural network mobilenet will reduce the classification accuracy, the hollow convolution kernel is introduced into a convolution layer of mobilenet model, and a S-Mobilenet model based on local receptive field expansion is proposed. The model is divided into three structures according to the location of the hole convolution kernel. Without increasing the number of parameters, it can expand the local receptive field of the convolution kernel and improve the classification accuracy. The experiment was carried out on caltech-101 data set, caltech-256 data set and animal classification database of Tubingen University. The results show that S-Mobilenet model can obtain better classification accuracy than Mobilenet, which can be improved by 2% at most.

A lightweight deep neural network with higher accuracy.

To improve accuracy of the MobileNet network, a new lightweight deep neural network is designed based on the MobileNetV2 network. Firstly, it modifies the network depth of MobileNetV2 to balance the image resolution, network width and depth to keep the gradient stable, which reduces the generation of gradient vanishing or gradient exploding. Secondly, it proposes an improved Bottleneck module by introducing channel attention mechanism. It assigns different weights for different channels according to the degree of relevance between the object features and channels. Therefore, the network can extract more effective features from a complex background. In the end, a new usage strategy of the improved Bottleneck is proposed. It uses the improved Bottleneck module in the second, fourth and fifth stages of MobileNetV2, and uses the original Bottleneck module in other states. Compared with MobileNetV2, MobileNetV3, ShuffleNetV2, GhostNet and HBONetmethods, the proposed method has the highest classification accuracy on the ImageNet-1K dataset, CIFAR-10 and CIFAR-100. Compared with YOLOV4-Lite methods based on these lightweight network networks, YOLOV4-Lite based on our proposed network also has the highest detection accuracy on the PASCAL VOC07+12 dataset.

A lightweight deep neural network for detection of mental states from physiological signals

A lightweight deep neural network for detection of mental states from physiological signals

LEQNet: Light Earthquake Deep Neural Network for Earthquake Detection and Phase Picking

Developing seismic signal detection and phase picking is an essential step for an on-site early earthquake warning system. A few deep learning approaches have been developed to improve the accuracy of seismic signal detection and phase picking. To run the existing deep learning models, high-throughput computing resources are required. In addition, the deep learning architecture must be optimized for mounting the model in small devices using low-cost sensors for earthquake detection. In this study, we designed a lightweight deep neural network model that operates on a very small device. We reduced the size of the deep learning model using the deeper bottleneck, recursive structure, and depthwise separable convolution. We evaluated our lightweight deep learning model using the Stanford Earthquake Dataset and compared it with EQTransformer. While our model size is reduced by 87.68% compared to EQTransformer, the performance of our model is comparable to that of EQTransformer.

A Novel Intrusion Detection Method Based on Lightweight Neural Network for Internet of Things

The purpose of a network intrusion detection (NID) is to detect intrusions in the network, which plays a critical role in ensuring the security of the Internet of Things (IoT). Recently, deep learning (DL) has achieved a great success in the field of intrusion detection. However, the limited computing capabilities and storage of IoT devices hinder the actual deployment of DL-based high-complexity models. In this article, we propose a novel NID method for IoT based on the lightweight deep neural network (LNN). In the data preprocessing stage, to avoid high-dimensional raw traffic features leading to high model complexity, we use the principal component analysis (PCA) algorithm to achieve feature dimensionality reduction. Besides, our classifier uses the expansion and compression structure, the inverse residual structure, and the channel shuffle operation to achieve effective feature extraction with low computational cost. For the multiclassification task, we adopt the NID loss that acts as a better loss function to replace the standard cross-entropy loss for dealing with the problem of uneven distribution of samples. The results of experiments on two real-world NID data sets demonstrate that our method has excellent classification performance with low model complexity and small model size, and it is suitable for classifying the IoT traffic of normal and attack scenarios.

LiteCovidNet: A lightweight deep neural network model for detection of COVID-19 using X-ray images.

The syndrome called COVID‐19 which was firstly spread in Wuhan, China has already been declared a globally “Pandemic.” To stymie the further spread of the virus at an early stage, detection needs to be done. Artificial Intelligence‐based deep learning models have gained much popularity in the detection of many diseases within the confines of biomedical sciences. In this paper, a deep neural network‐based “LiteCovidNet” model is proposed that detects COVID‐19 cases as the binary class (COVID‐19, Normal) and the multi‐class (COVID‐19, Normal, Pneumonia) bifurcated based on chest X‐ray images of the infected persons. An accuracy of 100% and 98.82% is achieved for binary and multi‐class classification respectively which is competitive performance as compared to the other recent related studies. Hence, our methodology can be used by health professionals to validate the detection of COVID‐19 infected patients at an early stage with convenient cost and better accuracy.

Lightweight single image super-resolution with attentive residual refinement network

In recent years, deep convolutional neural network (CNN) based single image super-resolution (SISR) methods have been demonstrated impressive performance in terms of quantitative metrics and visual effects. Most CNN-based SISR methods can learn the complex non-linear mapping between low-resolution (LR) images and their corresponding high-resolution (HR) images due to the powerful representation capabilities of deep convolutional neural networks. However, as the depth and width of the SISR networks increase, the parameters of SISR networks will increase dramatically, leading to huge computational cost and large memory consumption, making them impractical in real-world applications. To attack the above issues, we propose an accurate and lightweight deep convolutional neural network, named Attentive Residual Refinement Network (ARRFN), to recover the high-resolution image from the original low-resolution image directly for SISR. In general, our proposed ARRFN consists of three parts, a feature extraction block, a stack of attentive residual refinement blocks (ARRFB), and a multi-scale separable upscaling module (MSSU), respectively. Specifically, our ARRFB consists of two branches, a regular residual learning branch and an attentive residual refinement branch. The former conducts regular residual learning by two residual blocks while the latter refines the residual information from the two residual blocks of the former branch with an attentive residual mechanism to further enhance the representation capabilities of the network. Furthermore, a multi-scale separable upsampling module (MSSU) is proposed to replace the regular upsampling operation for better SR results. Extensive experiments on several standard benchmarks show that the proposed method outperforms state-of-the-art SR methods in terms of quantitative metrics, visual quality, memory footprint, and inference time.