Lightweight Deep Neural Network Research Articles

Ayur-PlantNet: An unbiased light weight deep convolutional neural network for Indian Ayurvedic plant species classification

Plant species classification using photo-based computer vision is challenging due to some crucial factors such as poor illumination, overlapping leaves, occlusions from cross-domain scene elements, inter-class similarities, intra-class variations, and blurred scene elements. The current state of art methods concerning photo-based plant species classification fails to deal with the challenges described above. In the proposed method, an unbiased lightweight deep convolutional neural network named Ayur-PlantNet is proposed to classify forty Ayurvedic plant species. The model built from scratch is trained and tested on 6000 samples with segmented plant regions. From a comparative study with pre-trained models; Resnet34, Resnet50, VGG16, MobileNetV3_Large, EfficientNetwork_B4, and Densenet121. It is noticed that the Ayur-PlantNet can produce an accuracy of 92.27% with reduced trainable parameters and computational complexity than pre-trained models. The experimental results prove that Ayur-PlantNet architecture is dominant compared to other deep learning models.

Violence Detection in Real-Life Audio Signals Using Lightweight Deep Neural Networks

The automatic detection of violent behavior in acoustic data has become an important research area due to its growing application potential in various surveillance and behavior monitoring tasks where it causes fewer privacy issues than video. In this paper, we propose two deep learning approaches of different characteristics to classify speech data containing violent behavior against data that only includes patterns of non-violent behavior. The first approach is based on conventional deep neural networks, while the second approach uses lightweight deep neural networks. Both utilize Mel-spectrogram images of speech signals as input to fine-tuned models. Our best lightweight model's classification accuracy is about 8% better than the previous state-of-the-art result using the same benchmark data. Lightweight models have fewer parameters and require fewer computing resources than conventional deep models. This can be an important advantage when being deployed on mobile or edge devices.

Classification Of Lung Cancer Using Lightweight Deep Neural Networks

Early recognition of lung nodules is critical for the effective conclusion and treatment of lung cancer. Numerous scientists have attempted various strategies, for example, thresholding, pattern recognition technique, computer-aided diagnosis, backpropagation calculation, etc. As of late, convolutional neural networks track down promising applications in numerous areas. In a conventional computer-aided detection (CAD) system, categorization of lung nodule detection is a challenging task. Each step of the lung nodule detection process depends on the classifier, which leads to poor recognition rate and a high false-positive rate. A classification approach for lung nodules based on deep hybrid learning is presented to address these challenges. This study aims to assess the performance of several low-memory, lightweight deep neural net (DNN) designs for image processing. The proposed lightweight deep neural model outperforms with an accuracy of 85.21% with reasonable specificity and sensitivity trade-off. The proposed work is based on binary classification networks such as vanilla 2D CNN, 2D SqueezeNet, and 2D MobileNet to recognize lung cancer in patient CT images with and without early-stage lung cancer.

Multi-Class Classification of Plant Leaf Diseases Using Feature Fusion of Deep Convolutional Neural Network and Local Binary Pattern

Plant diseases are one of the primary causes of decreased agricultural production quality and quantity. With ongoing changes in plant structure and cultivation techniques, new diseases are constantly arising on plant leaves. Thus, accurate classification and detection of plant leaf diseases in their early stages will limit the spread of the infection and support the healthy development of plant production. This work proposes a novel lightweight deep convolutional neural network (CNN) model for obtaining high-level hidden feature representations. The deep features are then fused with traditional handcrafted local binary pattern (LBP) features to capture local texture information in plant leaf images. The proposed model is trained and tested on three publicly available datasets (Apple Leaf, Tomato Leaf, and Grape Leaf). On the three datasets, the proposed approach achieves 99%, 96.6%, and 98.5% validation accuracies, respectively, and 98.8%, 96.5%, and 98.3% test accuracies, respectively. The results of the experiments show that the proposed approach can provide a better control solution for plant diseases.

Fully Automatic Left Ventricle Segmentation Using Bilateral Lightweight Deep Neural Network.

The segmentation of the left ventricle (LV) is one of the fundamental procedures that must be performed to obtain quantitative measures of the heart, such as its volume, area, and ejection fraction. In clinical practice, the delineation of LV is still often conducted semi-automatically, leaving it open to operator subjectivity. The automatic LV segmentation from echocardiography images is a challenging task due to poorly defined boundaries and operator dependency. Recent research has demonstrated that deep learning has the capability to employ the segmentation process automatically. However, the well-known state-of-the-art segmentation models still lack in terms of accuracy and speed. This study aims to develop a single-stage lightweight segmentation model that precisely and rapidly segments the LV from 2D echocardiography images. In this research, a backbone network is used to acquire both low-level and high-level features. Two parallel blocks, known as the spatial feature unit and the channel feature unit, are employed for the enhancement and improvement of these features. The refined features are merged by an integrated unit to segment the LV. The performance of the model and the time taken to segment the LV are compared to other established segmentation models, DeepLab, FCN, and Mask RCNN. The model achieved the highest values of the dice similarity index (0.9446), intersection over union (0.8445), and accuracy (0.9742). The evaluation metrics and processing time demonstrate that the proposed model not only provides superior quantitative results but also trains and segments the LV in less time, indicating its improved performance over competing segmentation models.

L-SecNet: Towards Secure and Lightweight Deep Neural Network Inference

The advances in machine learning technology has promoted its great potential for deep neural network (DNN) inference powered applications of Internet of Things (IoT), such as facial verification cameras and speech recognition assistants. The current deployment of these applications also raises serious privacy concerns, especially when sensitive individual information is accessed easily by various IoT devices. Fortunately, the cryptography-based solutions are able to execute secure inference without infringing the user’s raw data and the model owner’s proprietary model. However, existing works suffer from impractically high latency and low accuracy, stemming primarily from the evaluations of the non-linear layers in DNN. In this paper, we propose L-SecNet, a lightweight secure neural network inference system that provides efficient inference services without sacrificing accuracy and privacy. Specifically, to reduce latency caused by comparison operations in non-linear layers, we subtly combine additive secret sharing and multiplicative secret sharing to design a lightweight secure comparison protocol. Further, we approximate the commonly used and time-consuming activation functions (including Sigmoid and Tanh functions) with the non-linear sin function instead of the linear polynomial approximation functions in the existing works. In order to maintain low running latency while meeting the requirement of high accuracy, a secure and lightweight protocol for a sin function is proposed. Our theoretical analysis and empirical experiments evaluate the security and efficiency of the L-SecNet system. Compared with the state-of-the-art works, L-SecNet saves up to about 80 times bandwidth and about 53 times runtime.

Lightweight Deep Neural Networks for Ship Target Detection in SAR Imagery.

In recent years, deep convolutional neural networks (DCNNs) have been widely used in the task of ship target detection in synthetic aperture radar (SAR) imagery. However, the vast storage and computational cost of DCNN limits its application to spaceborne or airborne onboard devices with limited resources. In this paper, a set of lightweight detection networks for SAR ship target detection are proposed. To obtain these lightweight networks, this paper designs a network structure optimization algorithm based on the multi-objective firefly algorithm (termed NOFA). In our design, the NOFA algorithm encodes the filters of a well-performing ship target detection network into a list of probabilities, which will determine whether the lightweight network will inherit the corresponding filter structure and parameters. After that, the multi-objective firefly optimization algorithm (MFA) continuously optimizes the probability list and finally outputs a set of lightweight network encodings that can meet the different needs of the trade-off between detection network precision and size. Finally, the network pruning technology transforms the encoding that meets the task requirements into a lightweight ship target detection network. The experiments on SSDD and SDCD datasets prove that the method proposed in this paper can provide more flexible and lighter detection networks than traditional detection networks.

Electrocardiogram Signal Classification Using Lightweight DNN for Mobile Devices

Mobile users can use a mobile sensor to record ECG data, and on-device ECG classification can provide a more efficient diagnosis than standard care. Deep Neural Networks (DNNs) have excelled in artificial intelligence (AI)-based applications to extract patterns from the complex waveform to analyze the ECG waveform. Despite their superior performance, their high computational complexity severely limits their use in resource-constrained mobile phones. This study aimed to analyze various lightweight deep learning models for ECG classification on edge devices while maintaining comparable performance. Five DNNs, including SqueezeNet, ShuffleNet, MobileNet, EfficientNet, and NASNet, designed explicitly for resource-constrained mobile, are tested and compared for ECG classification. Time-series ECG signals are converted to a 2D scalogram using the continuous wavelet transform. These generated scalograms are classified into three categories: cardiac arrhythmia (ARR), congestive heart failure (CHF), and normal sinus rhythms (NSR). A comprehensive comparison is presented using accuracy analysis and the required trainable parameters among these models. The experimental analysis of PhysioNet's dataset shows that ShuffleNet has the fewest learnable parameters, with a validation accuracy of 87.50%. In contrast, SqueezeNet performed better, with 93.75% validation accuracy and 724K learnable parameters.

Developing a Deep Neural Network model for COVID-19 diagnosis based on CT scan images.

COVID-19 is most commonly diagnosed using a testing kit but chest X-rays and computed tomography (CT) scan images have a potential role in COVID-19 diagnosis. Currently, CT diagnosis systems based on Artificial intelligence (AI) models have been used in some countries. Previous research studies used complex neural networks, which led to difficulty in network training and high computation rates. Hence, in this study, we developed the 6-layer Deep Neural Network (DNN) model for COVID-19 diagnosis based on CT scan images. The proposed DNN model is generated to improve accurate diagnostics for classifying sick and healthy persons. Also, other classification models, such as decision trees, random forests and standard neural networks, have been investigated. One of the main contributions of this study is the use of the global feature extractor operator for feature extraction from the images. Furthermore, the 10-fold cross-validation technique is utilized for partitioning the data into training, testing and validation. During the DNN training, the model is generated without dropping out of neurons in the layers. The experimental results of the lightweight DNN model demonstrated that this model has the best accuracy of 96.71% compared to the previous classification models for COVID-19 diagnosis.

Time-Frequency Complexity Maps for EEG-Based Diagnosis of Alzheimer’s Disease Using a Lightweight Deep Neural Network

Alzheimer’s disease (AD) is a neurodegenerative disorder with an unknown etiology and a significant prevalence. Rapid and accurate detection of AD is crucial to assist in a more effective and tailored treatment plan to delay the progression of the disease. This paper introduces a novel approach based on a time-frequency complexity map (complextrogram) for the automated AD diagnosis. The complextrogram is the topographic complexity level of an EEG signal, plotted as a function of time and frequency. The complextrogram representations were fed into a well-known lightweight deep neural network called MobileNet for robust performance on resource and accuracy tradeoffs. The experiments were performed using a five-fold cross-validation technique on a publicly available database containing clinical EEG recordings from 24 patients with AD and 24 healthy, age-matched controls. The proposed pipeline provided competitive performance with just 2.2 M parameters and achieved the best overall accuracy for some locations in the frontal lobes (Fp2 and F8 channels). For both channels, the classification accuracy was 100%. Also, the violin plot was used to get further details of the distribution of complexity values for specific frequency rhythms. After statistical evaluation, it was observed that neurodegenerative conditions caused changes in chaotic behaviors, including increased delta complexity and decreased alpha complexity. Results demonstrated that the complextrogram representation proved its potency for the input quality required by the deep learning architectures. Furthermore, the complextrogram method is a promising pathway to discriminate and reflect the fundamental characteristics of AD abnormalities.

Contention alleviation in WiFi networks by using light-weight machine learning model

The random access MAC (e.g., CSMA/CA) is known to be vulnerable to heavy collisions when the transmission contention level is high. Heavy collisions have typically been mitigated by adjusting the length of the waiting interval imposed before individual transmission attempts, called ’backoff time’. For example, there exist a large body of studies to enhance the efficiency of the WiFi backoff mechanism. In this paper, we propose a novel scheme that inserts a configurable non-random duration, called offset, before the beginning of the conventional CSMA/CA backoff. Since the existing backoff mechanism is untouched, the proposed scheme is compatible with all legacy and future WiFi standards. The size of offset is dynamically determined at run time by using a lightweight machine learning model. While machine learning models have been applied to the problem of contention alleviation in WiFi networks, the proposed scheme is unique in its flexibility and low overhead for model training and inferencing. Our scheme employs a simple lightweight DNN (Deep Neural Network) model which does not require a large amount of training data. The complexity of our machine learning model does not increase even if the number of the devices in the WiFi networks increases. Furthermore, the proposed scheme is executed in a distributed fashion, so that (near) real-time adaptation to dynamic change of traffic condition is possible at each end device that is not equipped with high computing resource (i.e., on-device DNN model execution).

Light-Weight Deep Learning Techniques with Advanced Processing for Real-Time Hand Gesture Recognition

In the discipline of hand gesture and dynamic sign language recognition, deep learning approaches with high computational complexity and a wide range of parameters have been an extremely remarkable success. However, the implementation of sign language recognition applications for mobile phones with restricted storage and computing capacities is usually greatly constrained by those limited resources. In light of this situation, we suggest lightweight deep neural networks with advanced processing for real-time dynamic sign language recognition (DSLR). This paper presents a DSLR application to minimize the gap between hearing-impaired communities and regular society. The DSLR application was developed using two robust deep learning models, the GRU and the 1D CNN, combined with the MediaPipe framework. In this paper, the authors implement advanced processes to solve most of the DSLR problems, especially in real-time detection, e.g., differences in depth and location. The solution method consists of three main parts. First, the input dataset is preprocessed with our algorithm to standardize the number of frames. Then, the MediaPipe framework extracts hands and poses landmarks (features) to detect and locate them. Finally, the features of the models are passed after processing the unification of the depth and location of the body to recognize the DSL accurately. To accomplish this, the authors built a new American video-based sign dataset and named it DSL-46. DSL-46 contains 46 daily used signs that were presented with all the needed details and properties for recording the new dataset. The results of the experiments show that the presented solution method can recognize dynamic signs extremely fast and accurately, even in real-time detection. The DSLR reaches an accuracy of 98.8%, 99.84%, and 88.40% on the DSL-46, LSA64, and LIBRAS-BSL datasets, respectively.

Smart farming for detection and identification of tomato plant diseases using light weight deep neural network

Smart farming for detection and identification of tomato plant diseases using light weight deep neural network

Union-net: lightweight deep neural network model suitable for small data sets

Traditional deep learning models prefer large data sets, and in reality small data sets are easier to obtain. It is more practical to build models suitable for small data sets. Based on CNN, this paper proposes the concept of union convolution to build a deep learning model Union-net that is suitable for small data sets. The Union-net has small model size and superior performance. In this paper, the model is tested based on multiple commonly used data sets. The experimental results show that Union-net outperforms most models when dealing with small datasets, and Union-net outperforms other models when dealing with complex classification tasks or dealing with few-shot datasets. The codes for this paper have been uploaded to https://github.com/yeaso/union-net .

A monocular visual SLAM system augmented by lightweight deep local feature extractor using in-house and low-cost LIDAR-camera integrated device

ABSTRACT Simultaneous Localization and Mapping (SLAM) has been widely used in emergency response, self-driving and city-scale 3D mapping and navigation. Recent deep-learning based feature point extractors have demonstrated superior performance in dealing with the complex environmental challenges (e.g. extreme lighting) while the traditional extractors are struggling. In this paper, we have successfully improved the robustness and accuracy of a monocular visual SLAM system under various complex scenes by adding a deep learning based visual localization thread as an augmentation to the visual SLAM framework. In this thread, our feature extractor with an efficient lightweight deep neural network is used for absolute pose and scale estimation in real time using the highly accurate georeferenced prior map database at 20cm geometric accuracy created by our in-house and low-cost LiDAR and camera integrated device. The closed-loop error provided by our SLAM system with and without this enhancement is 1.03m and 18.28m respectively. The scale estimation of the monocular visual SLAM is also significantly improved (0.01 versus 0.98). In addition, a novel camera-LiDAR calibration workflow is also provided for large-scale 3D mapping. This paper demonstrates the application and research potential of deep-learning based vision SLAM with image and LiDAR sensors.

Lightweight Deep Neural Network for Joint Learning of Underwater Object Detection and Color Conversion.

Underwater image processing has been shown to exhibit significant potential for exploring underwater environments. It has been applied to a wide variety of fields, such as underwater terrain scanning and autonomous underwater vehicles (AUVs)-driven applications, such as image-based underwater object detection. However, underwater images often suffer from degeneration due to attenuation, color distortion, and noise from artificial lighting sources as well as the effects of possibly low-end optical imaging devices. Thus, object detection performance would be degraded accordingly. To tackle this problem, in this article, a lightweight deep underwater object detection network is proposed. The key is to present a deep model for jointly learning color conversion and object detection for underwater images. The image color conversion module aims at transforming color images to the corresponding grayscale images to solve the problem of underwater color absorption to enhance the object detection performance with lower computational complexity. The presented experimental results with our implementation on the Raspberry pi platform have justified the effectiveness of the proposed lightweight jointly learning model for underwater object detection compared with the state-of-the-art approaches.

Improving QoE of Deep Neural Network Inference on Edge Devices: A Bandit Approach

Edge devices, including, in particular, mobile devices, have been emerging as an increasingly more important platform for deep neural network (DNN) inference. Typically, multiple lightweight DNN models generated using different architectures and/or compression schemes can fit into a device, thus selecting an optimal one is crucial in order to maximize the users’ Quality of Experience (QoE) for edge inference. The existing approaches to device-aware DNN optimization are usually time consuming and not scalable in view of extremely diverse edge devices. More importantly, they focus on optimizing standard performance metrics (e.g., accuracy and latency), which may not translate into improvement of the users’ actual subjective QoE. In this article, we propose a novel automated and user-centric DNN selection engine, called <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\mathsf {Aquaman}$ </tex-math></inline-formula> , which keeps users into a closed loop and leverages their QoE feedback to guide DNN selection decisions. The core of <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\mathsf {Aquaman}$ </tex-math></inline-formula> is a neural network-based QoE predictor, which is continuously updated online. Additionally, we use neural bandit learning to balance exploitation and exploration, with a provably efficient QoE performance. Finally, we evaluate <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\mathsf {Aquaman}$ </tex-math></inline-formula> on a 15-user experimental study as well as synthetic simulations, demonstrating the effectiveness of <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\mathsf {Aquaman}$ </tex-math></inline-formula> .

An Explainable and Lightweight Deep Convolutional Neural Network for Quality Detection of Green Coffee Beans

In recent years, the demand for coffee has increased tremendously. During the production process, green coffee beans are traditionally screened manually for defective beans before they are packed into coffee bean packages; however, this method is not only time-consuming but also increases the rate of human error due to fatigue. Therefore, this paper proposed a lightweight deep convolutional neural network (LDCNN) for a quality detection system of green coffee beans, which combined depthwise separable convolution (DSC), squeeze-and-excite block (SE block), skip block, and other frameworks. To avoid the influence of low parameters of the lightweight model caused by the model training process, rectified Adam (RA), lookahead (LA), and gradient centralization (GC) were included to improve efficiency; the model was also put into the embedded system. Finally, the local interpretable model-agnostic explanations (LIME) model was employed to explain the predictions of the model. The experimental results indicated that the accuracy rate of the model could reach up to 98.38% and the F1 score could be as high as 98.24% when detecting the quality of green coffee beans. Hence, it can obtain higher accuracy, lower computing time, and lower parameters. Moreover, the interpretable model verified that the lightweight model in this work was reliable, providing the basis for screening personnel to understand the judgment through its interpretability, thereby improving the classification and prediction of the model.

A new lightweight deep neural network for surface scratch detection.

This paper aims to develop a lightweight convolutional neural network, WearNet, to realise automatic scratch detection for components in contact sliding such as those in metal forming. To this end, a large surface scratch dataset obtained from cylinder-on-flat sliding tests was used to train the WearNet with appropriate training parameters such as learning rate, gradient algorithm and mini-batch size. A comprehensive investigation on the network response and decision mechanism was also conducted to show the capability of the developed WearNet. It was found that compared with the existing networks, WearNet can realise an excellent classification accuracy of 94.16% with a much smaller model size and faster detection speed. Besides, WearNet outperformed other state-of-the-art networks when a public image database was used for network evaluation. The application of WearNet in an embedded system further demonstrated such advantages in the detection of surface scratches in sheet metal forming processes.

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.