Lightweight Network Architecture Research Articles

DeformableFishNet: a high-precision lightweight target detector for underwater fish identification

The application of computer vision in fish identification facilitates researchers and managers to better comprehend and safeguard the aquatic ecological environment. Numerous researchers have harnessed deep learning methodologies for studying fish species identification. Nonetheless, this endeavor still encounters challenges such as high computational costs, a substantial number of parameters, and limited practicality. To address these issues, we propose a lightweight network architecture incorporating deformable convolutions, termed DeformableFishNet. Within DeformableFishNet, an efficient global coordinate attention module (EGCA) is introduced alongside a deformable convolution network (EDCN/EC2f), which is grounded in EGCA, to tackle the deformation of fish bodies induced by swimming motions. Additionally, an EC2f-based feature pyramid network (EDBFPN) and an efficient multi-scale decoupling head (EMSD Head) are proposed to extract multi-scale fish features within a lightweight framework. DeformableFishNet was deployed on our freshwater fish dataset, with experimental outcomes illustrating its efficacy, achieving a mean average precision (mAP) of 96.3%. The model comprises 1.7 million parameters and entails 4.7 billion floating-point operations (FLOPs). Furthermore, we validated DeformableFishNet on three public underwater datasets, yielding respective mAPs of 98%, 99.4%, and 83.6%. The experiments show that DeformableFishNet is suitable for underwater identification of various scenes.

DeformableFishNet: a high-precision lightweight target detector for underwater fish identification

The application of computer vision in fish identification facilitates researchers and managers to better comprehend and safeguard the aquatic ecological environment. Numerous researchers have harnessed deep learning methodologies for studying fish species identification. Nonetheless, this endeavor still encounters challenges such as high computational costs, a substantial number of parameters, and limited practicality. To address these issues, we propose a lightweight network architecture incorporating deformable convolutions, termed DeformableFishNet. Within DeformableFishNet, an efficient global coordinate attention module (EGCA) is introduced alongside a deformable convolution network (EDCN/EC2f), which is grounded in EGCA, to tackle the deformation of fish bodies induced by swimming motions. Additionally, an EC2f-based feature pyramid network (EDBFPN) and an efficient multi-scale decoupling head (EMSD Head) are proposed to extract multi-scale fish features within a lightweight framework. DeformableFishNet was deployed on our freshwater fish dataset, with experimental outcomes illustrating its efficacy, achieving a mean average precision (mAP) of 96.3%. The model comprises 1.7 million parameters and entails 4.7 billion floating-point operations (FLOPs). Furthermore, we validated DeformableFishNet on three public underwater datasets, yielding respective mAPs of 98%, 99.4%, and 83.6%. The experiments show that DeformableFishNet is suitable for underwater identification of various scenes.

Efficient and lightweight convolutional neural network architecture search methods for object classification

Determining the architecture of deep learning models is a complex task. Several automated search techniques have been proposed, but these methods typically require high-performance graphics processing units (GPUs), manual parameter adjustments, and specific training approaches. This study introduces an efficient, lightweight convolutional neural network architecture search approach tailored for object classification. It features an optimized search space design and a novel controller design.This study introduces a refined search space design incorporating optimizations in both spatial and operational aspects. The focus is on the synergistic integration of convolutional units, dimension reduction units, and the stacking of Convolutional Neural Network (CNN) architectures. To enhance the search space, ShuffleNet modules are integrated, reducing the number of parameters and training time. Additionally, BlurPool is implemented in the dimension reduction unit operation to achieve translational invariance, alleviate the gradient vanishing problem, and optimize unit compositions. Moreover, an innovative controller model, Stage LSTM, is proposed based on Long Short-Term Memory (LSTM) to generate lightweight architectural sequences. In conclusion, the refined search space design and the Stage LSTM controller model are synergistically combined to establish an efficient and lightweight architecture search technique termed Stage and Lightweight Network Architecture Search (SLNAS).The experimental results highlight the superior performance of the optimized search space design, primarily when implemented with the Stage LSTM controller model. This approach shows significantly improved accuracy and stability compared to random, traditional LSTM, and Genetic Algorithm (GA) controller models, with statistically significant differences. Notably, the Stage LSTM controller excels in accuracy while producing models with fewer parameters within the expanded architecture search space.The study adopts the Stage LSTM controller model due to its ability to approximate optimal sequence structures, particularly when combined with the optimized search space design, referred to as SLNAS. SLNAS's performance is evaluated through experiments and comparisons with other Neural Architecture Search (NAS) and object classification methods from different researchers. These experiments consider model parameters, hardware resources, model stability, and multiple datasets. The results show that SLNAS achieves a low error rate of 2.86 % on the CIFAR-10 dataset after just 0.2 days of architecture search, matching the performance of manually designed models but using only 2 % of the parameters. SLNAS consistently demonstrates robust performance across various image classification domains, with an approximate parameter count 700,000.To summarize, SLNAS emerges as a highly effective automated network architecture search method tailored for image classification. It streamlines the model design process, making it accessible to researchers without specialized knowledge in deep learning. Optimizing this method unlocks the full potential of deep learning across diverse research areas. Interested parties can publicly access the source code and pre-trained models through the following link: https://github.com/huanyu-chen/LNASG-and-SLNAS-model.

A lightweight network architecture for traffic sign recognition based on enhanced LeNet-5 network.

As an important part of the unmanned driving system, the detection and recognition of traffic sign need to have the characteristics of excellent recognition accuracy, fast execution speed and easy deployment. Researchers have applied the techniques of machine learning, deep learning and image processing to traffic sign recognition successfully. Considering the hardware conditions of the terminal equipment in the unmanned driving system, in this research work, the goal was to achieve a convolutional neural network (CNN) architecture that is lightweight and easily implemented for an embedded application and with excellent recognition accuracy and execution speed. As a classical CNN architecture, LeNet-5 network model was chosen to be improved, including image preprocessing, improving spatial pool convolutional neural network, optimizing neurons, optimizing activation function, etc. The test experiment of the improved network architecture was carried out on German Traffic Sign Recognition Benchmark (GTSRB) database. The experimental results show that the improved network architecture can obtain higher recognition accuracy in a short interference time, and the algorithm loss is significantly reduced with the progress of training. At the same time, compared with other lightweight network models, this network architecture gives a good recognition result, with a recognition accuracy of 97.53%. The network structure is simple, the algorithm complexity is low, and it is suitable for all kinds of terminal equipment, which can have a wider application in unmanned driving system.

Semi-TSGAN: Semi-Supervised Learning for Highlight Removal Based on Teacher-Student Generative Adversarial Network.

Despite recent notable advancements in highlight image restoration techniques, the dearth of annotated data and the lightweight deployment of highlight removal networks pose significant impediments to further advancements in the field. In this paper, to the best of our knowledge, we first propose a semi-supervised learning paradigm for highlight removal, merging the fusion version of a teacher-student model and a generative adversarial network, featuring a lightweight network architecture. Initially, we establish a dependable repository to house optimal predictions as pseudo ground truth through empirical analyses guided by the most reliable No-Reference Image Quality Assessment (NR-IQA) method. This method serves to assess rigorously the quality of model predictions. Subsequently, addressing concerns regarding confirmation bias, we integrate contrastive regularization into the framework to curtail the risk of overfitting on inaccurate labels. Finally, we introduce a comprehensive feature aggregation module and an extensive attention mechanism within the generative network, considering a balance between network performance and computational efficiency. Our experimental evaluations encompass comprehensive assessments on both full-reference and non-reference highlight benchmarks. The results demonstrate conclusively the substantive quantitative and qualitative enhancements achieved by our proposed algorithm in comparison to state-of-the-art methodologies.

AVD-YOLOv5: a new lightweight network architecture for high-speed aortic valve detection from a new and large echocardiography dataset.

Heart disease detection is currently gaining widespread attention as a means to enhance the accuracy of cardiologists' diagnoses from cardiac images and reduce diagnosis time. Although high-resolution computed tomography (CT) images are typically favored for heart disease detection, the drawbacks of cost and radiation exposure to patients necessitate alternative approaches. In this context, utilizing ultrasound images becomes pivotal to mitigate radiation risks and maintain cost-effectiveness. In this paper, we propose a novel lightweight model, AVD-YOLOv5, designed for automated aortic valve detection on echocardiography images. This model incorporates several enhancements to the YOLOv5 architecture. Notably, the depth-wise separable convolution significantly contributes to the model's lightweight design by reducing the number of parameters while maintaining precision. We have also created a new and larger dataset comprising 260 echocardiography images specifically for aortic valve detection. Experimental results indicate that the precision value of the modified ADV-YOLOv5 model stands at 94.3%, with a recall value of 86.8%. The model also demonstrates a notable 67% reduction in inference time compared to the original YOLOv5 model. Although there is a marginal reduction in precision by 0.94%, the model's efficiency is significantly increased. The proposed system can be used by cardiologists for more efficient and reliable diagnosis.

A Zero-Reference Low-Light Image-Enhancement Approach Based on Noise Estimation

A novel zero-reference low-light image-enhancement approach based on noise estimation (ZLEN) is proposed to mitigate noise interference in image-enhancement processes, while the tenets of zero-reference and lightweight network architecture are maintained. ZLEN improves the high-order curve expression governing the mapping of low-light images to their enhanced counterparts, addressing image noise through a meticulously designed noise-estimation module and a zero-reference noise loss function. First, the higher-order curve expression with a noise term is defined, and then the noise map undergoes feature extraction through the semantic-aware attention module; following this, the resulting features are integrated with the low-light image. Ultimately, a lightweight convolutional neural network is adjusted to estimate higher-order curve parameters that link the low-light image to its enhanced version. Notably, ZLEN achieves luminance enhancement and noise reduction without paired or unpaired training data. Rigorous qualitative and quantitative evaluations were conducted on diverse benchmark datasets, demonstrating that ZLEN attained state-of-the-art (SOAT) status among existing zero-reference and unpaired-reference image-enhancement methodologies, while it exhibited comparable performance to full-reference image-enhancement methods. To confirm the practicality and robustness of ZLEN, the luminance enhancement was applied to mine images, which yielded satisfactory results.

Real-Time Traffic Light Recognition with Lightweight State Recognition and Ratio-Preserving Zero Padding

As online shopping is becoming mainstream, driven by the social impact of Coronavirus disease-2019 (COVID-19) as well as the development of Internet services, the demand for autonomous delivery mobile robots is rapidly increasing. This trend has brought the autonomous mobile robot market to a new turning point, with expectations that numerous mobile robots will be driving on roads with traffic. To achieve these expectations, autonomous mobile robots should precisely perceive the situation on roads with traffic. In this paper, we revisit and implement a real-time traffic light recognition system with a proposed lightweight state recognition network and ratio-preserving zero padding, which is a two-stage system consisting of a traffic light detection (TLD) module and a traffic light status recognition (TLSR) module. For the TLSR module, this work proposes a lightweight state recognition network with a small number of weight parameters, because the TLD module needs more weight parameters to find the exact location of traffic lights. Then, the proposed effective and lightweight network architecture is constructed by using skip connection, multifeature maps with different sizes, and kernels of appropriately tuned sizes. Therefore, the network has a negligible impact on the overall processing time and minimal weight parameters while maintaining high performance. We also propose to utilize a ratio-preserving zero padding method for data preprocessing for the TLSR module to enhance recognition accuracy. For the TLD module, extensive evaluations with varying input sizes and backbone network types are conducted, and then appropriate values for those factors are determined, which strikes a balance between detection performance and processing time. Finally, we demonstrate that our traffic light recognition system, utilizing the TLD module’s determined parameters, the proposed network architecture for the TLSR module, and the ratio-preserving zero padding method can reliably detect the location and state of traffic lights in real-world videos recorded in Gumi and Deagu, Korea, while maintaining at least 30 frames per second for real-time operation.

Lightweight air-to-air unmanned aerial vehicle target detection model

The rapid expansion of the drone industry has resulted in a substantial increase in the number of low-altitude drones, giving rise to concerns regarding collision avoidance and countermeasure strategies among these unmanned aerial vehicles. These challenges underscore the urgent need for air-to-air drone target detection. An effective target detection model must exhibit high accuracy, real-time capabilities, and a lightweight network architecture to achieve a balance between precision and speed when deployed on embedded devices. In response to these requirements, we initially curated a dataset comprising over 10,000 images of low-altitude operating drones. This dataset encompasses diverse and intricate backgrounds, significantly enhancing the model’s training capacity. Subsequently, a series of enhancements were applied to the YOLOv5 algorithm to realize lightweight object detection. A novel feature extraction network, CF2-MC, streamlined the feature extraction process, while an innovative module, MG, in the feature fusion section aimed to improve detection accuracy and reduce model complexity. Concurrently, the original CIoU loss function was replaced with the EIoU loss function to further augment the model’s accuracy. Experimental results demonstrate an enhancement in the accuracy of drone target detection, achieving mAP values of 95.4% on the UAVfly dataset and 82.2% on the Det-Fly dataset. Finally, real-world testing conducted on the Jetson TX2 revealed that the YOLOv5s-ngn model achieved an average inference speed of 14.5 milliseconds per image. The code utilized in this paper can be accessed via https://github.com/lucien22588/yolov5-ngn.git.

6DoF-3D: Efficient and accurate 3D object detection using six degrees-of-freedom for autonomous driving

With the advancement of automobile industry, autonomous driving has become an integral component of contemporary vehicles. In this context, the 3D spatial perception of vehicles plays a crucial role. The complexity of 3D object detection dramatically increased due to the explosion of degrees-of-freedom for 3D bounding boxes. Therefore, current 3D detection algorithms struggle to achieve high accuracy and efficiency simultaneously. To address this problem, we introduce a single-stage, LiDAR-based, and efficiency–accuracy balanced 3D object detector named 6DoF-3D. The proposed network architecture comprises a backbone of CSPDarknet53, a multi-scale feature fusion neck, and a detection head. We transform 3D point clouds into pseudo-images, with critical spatial features retained. 6DoF-3D presents a novel encoding strategy to prune the redundant degree-of-freedom in 3D detection, resulting in a lightweight network architecture. We put forward a 2.5D Euler-Region-Proposal network for precise classification and regression of 2.5D bounding boxes. The pruned dimension is compensated with a geometric transformation decoder. Extensive experiments conducted on the KITTI dataset demonstrate that 6DoF-3D achieves state-of-the-art accuracy of 89.62% for 3D car detection at the moderate level. Additionally, the mean average precision reaches 66.72% at the while maintaining a real-time performance of 33.21 FPS for cars, pedestrians, and cyclists.

Focus-RCNet: a lightweight recyclable waste classification algorithm based on focus and knowledge distillation

Waste pollution is a significant environmental problem worldwide. With the continuous improvement in the living standards of the population and increasing richness of the consumption structure, the amount of domestic waste generated has increased dramatically, and there is an urgent need for further treatment. The rapid development of artificial intelligence has provided an effective solution for automated waste classification. However, the high computational power and complexity of algorithms make convolutional neural networks unsuitable for real-time embedded applications. In this paper, we propose a lightweight network architecture called Focus-RCNet, designed with reference to the sandglass structure of MobileNetV2, which uses deeply separable convolution to extract features from images. The Focus module is introduced to the field of recyclable waste image classification to reduce the dimensionality of features while retaining relevant information. To make the model focus more on waste image features while keeping the number of parameters small, we introduce the SimAM attention mechanism. In addition, knowledge distillation was used to further compress the number of parameters in the model. By training and testing on the TrashNet dataset, the Focus-RCNet model not only achieved an accuracy of 92% but also showed high deployment mobility.

Learning Template-Constraint Real-Time Siamese Tracker for Drone AI Devices via Concatenation

Significant progress has been made in object tracking tasks thanks to the application of deep learning. However, current deep neural network-based object tracking methods often rely on stacking sub-modules and introducing complex structures to improve tracking accuracy. Unfortunately, these approaches are inefficient and limit the feasibility of deploying efficient trackers on drone AI devices. To address these challenges, this paper introduces ConcatTrk, a high-speed object tracking method designed specifically for drone AI devices. ConcatTrk utilizes a lightweight network architecture, enabling real-time tracking on edge devices. Specifically, the proposed method primarily uses the concatenation operation to construct its core tracking steps, including multi-scale feature fusion, intra-frame feature matching, and dynamic template updating, which aim to reduce the computational overhead of the tracker. To ensure tracking performance in UAV tracking scenarios, ConcatTrk implements a learnable feature matching operator along with a simple and efficient template constraint branch, which enables accurate tracking by discriminatively matching features and incorporating periodic template updates. Results of comprehensive experiments on popular benchmarks, including UAV123, OTB100, and LaSOT, show that ConcatTrk has achieved promising accuracy and attained a tracking speed of 41 FPS on an edge AI device, Nvidia AGX Xavier. ConcatTrk runs 8× faster than the SOTA tracker TransT while using 4.9× fewer FLOPs. Real-world tests on the drone platform have strongly validated its practicability, including real-time tracking speed, reliable accuracy, and low power consumption.

Real-Time Segmentation of Unstructured Environments by Combining Domain Generalization and Attention Mechanisms.

This paper presents a focused investigation into real-time segmentation in unstructured environments, a crucial aspect for enabling autonomous navigation in off-road robots. To address this challenge, an improved variant of the DDRNet23-slim model is proposed, which includes a lightweight network architecture and reclassifies ten different categories, including drivable roads, trees, high vegetation, obstacles, and buildings, based on the RUGD dataset. The model's design includes the integration of the semantic-aware normalization and semantic-aware whitening (SAN-SAW) module into the main network to improve generalization ability beyond the visible domain. The model's segmentation accuracy is improved through the fusion of channel attention and spatial attention mechanisms in the low-resolution branch to enhance its ability to capture fine details in complex scenes. Additionally, to tackle the issue of category imbalance in unstructured scene datasets, a rare class sampling strategy (RCS) is employed to mitigate the negative impact of low segmentation accuracy for rare classes on the overall performance of the model. Experimental results demonstrate that the improved model achieves a significant 14% increase mIoU in the invisible domain, indicating its strong generalization ability. With a parameter count of only 5.79M, the model achieves mAcc of 85.21% and mIoU of 77.75%. The model has been successfully deployed on a a Jetson Xavier NX ROS robot and tested in both real and simulated orchard environments. Speed optimization using TensorRT increased the segmentation speed to 30.17 FPS. The proposed model strikes a desirable balance between inference speed and accuracy and has good domain migration ability, making it applicable in various domains such as forestry rescue and intelligent agricultural orchard harvesting.

Transfer learning based cascaded deep learning network and mask recognition for COVID-19.

The COVID-19 is still spreading today, and it has caused great harm to human beings. The system at the entrance of public places such as shopping malls and stations should check whether pedestrians are wearing masks. However, pedestrians often pass the system inspection by wearing cotton masks, scarves, etc. Therefore, the detection system not only needs to check whether pedestrians are wearing masks, but also needs to detect the type of masks. Based on the lightweight network architecture MobilenetV3, this paper proposes a cascaded deep learning network based on transfer learning, and then designs a mask recognition system based on the cascaded deep learning network. By modifying the activation function of the MobilenetV3 output layer and the structure of the model, two MobilenetV3 networks suitable for cascading are obtained. By introducing transfer learning into the training process of two modified MobilenetV3 networks and a multi-task convolutional neural network, the ImagNet underlying parameters of the network models are obtained in advance, which reduces the computational load of the models. The cascaded deep learning network consists of a multi-task convolutional neural network cascaded with these two modified MobilenetV3 networks. A multi-task convolutional neural network is used to detect faces in images, and two modified MobilenetV3 networks are used as the backbone network to extract the features of masks. After comparing with the classification results of the modified MobilenetV3 neural network before cascading, the classification accuracy of the cascading learning network is improved by 7%, and the excellent performance of the cascading network can be seen.

Real-time defect detection of TFT-LCD displays using a lightweight network architecture

Real-time defect detection of TFT-LCD displays using a lightweight network architecture

DEFAEK: Domain Effective Fast Adaptive Network for Face Anti-Spoofing

Existing deep learning based face anti-spoofing (FAS) or deepfake detection approaches usually rely on large-scale datasets and powerful networks with significant amount of parameters to achieve satisfactory performance. However, these make them resource-heavy and unsuitable for handheld devices. Moreover, they are limited by the types of spoof in the dataset they train on and require considerable training time. To produce a robust FAS model, they need large datasets covering the widest variety of predefined presentation attacks possible. Testing on new or unseen attacks or environments generally results in poor performance. Ideally, the FAS model should learn discriminative features that can generalize well even on unseen spoof types. In this paper, we propose a fast learning approach called Domain Effective Fast Adaptive nEt-worK (DEFAEK), a face anti-spoofing approach based on the optimization-based meta-learning paradigm that effectively and quickly adapts to new tasks. DEFAEK treats differences in an environment as domains and simulates multiple domain shifts during training. To further improve the effectiveness and efficiency of meta-learning, we adopt the metric learning in the inner loop update with careful sample selection. With extensive experiments on the challenging CelebA-Spoof and FaceForensics++ datasets, the evaluation results show that DEFAEK can learn cues independent of the environment with good generalization capability. In addition, the resulting model is lightweight following the design principle of modern lightweight network architecture and still generalizes well on unseen classes. In addition, we also demonstrate our model’s capabilities by comparing the numbers of parameters, FLOPS, and model performance with other state-of-the-art methods.

Detection and Recognition Algorithm of Arbitrary-Oriented Oil Replenishment Target in Remote Sensing Image

In view of the fact that the aerial images of UAVs are usually taken from a top-down perspective, there are large changes in spatial resolution and small targets to be detected, and the detection method of natural scenes is not effective in detecting under the arbitrary arrangement of remote sensing image direction, which is difficult to apply to the detection demand scenario of road technology status assessment, this paper proposes a lightweight network architecture algorithm based on MobileNetv3-YOLOv5s (MR-YOLO). First, the MobileNetv3 structure is introduced to replace part of the backbone network of YOLOv5s for feature extraction so as to reduce the network model size and computation and improve the detection speed of the target; meanwhile, the CSPNet cross-stage local network is introduced to ensure the accuracy while reducing the computation. The focal loss function is improved to improve the localization accuracy while increasing the speed of the bounding box regression. Finally, by improving the YOLOv5 target detection network from the prior frame design and the bounding box regression formula, the rotation angle method is added to make it suitable for the detection demand scenario of road technology status assessment. After a large number of algorithm comparisons and data ablation experiments, the feasibility of the algorithm was verified on the Xinjiang Altay highway dataset, and the accuracy of the MR-YOLO algorithm was as high as 91.1%, the average accuracy was as high as 92.4%, and the detection speed reached 96.8 FPS. Compared with YOLOv5s, the p-value and mAP values of the proposed algorithm were effectively improved. It can be seen that the proposed algorithm improves the detection accuracy and detection speed while greatly reducing the number of model parameters and computation.

A Sparsified Densely Connected Network with Separable Convolution for Finger-Vein Recognition

At present, ResNet and DenseNet have achieved significant performance gains in the field of finger-vein biometric recognition, which is partially attributed to the dominant design of cross-layer skip connection. In this manner, features from multiple layers can be effectively aggregated to provide sufficient discriminant representation. Nevertheless, an over-dense connection pattern may induce channel expansion of feature maps and excessive memory consumption. To address these issues, we proposed a low memory overhead and fairly lightweight network architecture for finger-vein recognition. The core components of the proposed network are a sequence of sparsified densely connected blocks with symmetric structure. In each block, a novel connection cropping strategy is adopted to balance the channel ratio of input/output feature maps. Beyond this, to facilitate smaller model volume and faster convergence, we substitute the standard convolutional kernels with separable convolutional kernels and introduce a robust loss metric that is defined on the geodesic distance of angular space. Our proposed sparsified densely connected network with separable convolution (hereinafter dubbed ‘SC-SDCN’) has been tested on two benchmark finger-vein datasets, including the Multimedia Lab of Chonbuk National University (MMCBNU)and Finger Vein of Universiti Sains Malaysia (FV-USM), and the advantages of our SC-SDCN can be evident from the experimental results. Specifically, an equal error rate (EER) of 0.01% and an accuracy of 99.98% are obtained on the MMCBNU dataset, and an EER of 0.45% and an accuracy of 99.74% are obtained on the FV-USM dataset.

Lightweight network learning with Zero-Shot Neural Architecture Search for UAV images

Lightweight Network Architecture is essential for autonomous and intelligent monitoring of Unmanned Aerial Vehicles (UAVs), such as in object detection, image segmentation, and crowd counting applications. The state-of-the-art lightweight network learning based on Neural Architecture Search (NAS) usually costs enormous computation resources. Alternatively, low-performance embedded platforms and high-resolution drone images pose a challenge for lightweight network learning. To alleviate this problem, this paper proposes a new lightweight object detection model, called GhostShuffleNet (GSNet), for UAV images, which is built based on Zero-Shot Neural Architecture Search. This paper also introduces the new components which compose GSNet, namely GhostShuffle units (loosely based on ShuffleNetV2) and the backbone GSmodel-L. Firstly, a lightweight search space is constructed with the GhostShuffle (GS) units to reduce the parameters and floating-point operations (FLOPs). Secondly, the parameters, FLOPs, layers, and memory access cost (MAC) as constraints add to search strategy on a Zero-Shot Neural structure search algorithm, which then searches for an optimal network GSmodel-L. Finally, the optimal GSmodel-L is used as the backbone network and a Ghost-PAN feature fusion module and detection heads are added to complete the design of the lightweight object detection network (GSNet). Extensive experiments are conducted on the VisDrone2019 (14.92%mAP) dataset and the our UAV-OUC-DET (8.38%mAP) dataset demonstrating the efficiency and effectiveness of GSNet. The completed code is available at: https://github.com/yfq-yy/GSNet.

Construction of Edge Computing Platform Using 3D LiDAR and Camera Heterogeneous Sensing Fusion for Front Obstacle Recognition and Distance Measurement System

This research aims to utilise heterogeneous sensor fusion using 3D Light Detection and Ranging (LiDAR) and cameras, combined with an object recognition system and a ranging system, to construct an edge computing platform such that a vehicle equipped with the platform can perform computations offline in real time. This work comprises two main sections: the first is heterogeneous fusion, and the second is obstacle recognition and ranging detection. To achieve heterogeneous sensor fusion, 3D–3D point matching was used to find rigid body transformation between two sensors and finally project the LiDAR 3D point cloud image onto the 2D image. For object recognition, YOLOv4-Tiny was used as the detection network. A lightweight network architecture and high computational speed could be effectively used on edge computing hardware with limited performance. Further, by drawing the bounding box, we could detect the point cloud within the bounding box to estimate the distance to the obstacle. For detecting distance, we conducted experiments in two ways: ‘minimum point in box’ and ‘median point in box’ and compared the results. With heterogeneous sensor fusion, object recognition and the ranging system, detecting the category and distance of obstacles ahead of the vehicle was possible in real time. Furthermore, integrating the edge computing platform architecture enabled moving the entire system offline, making it an independent system that returns results in real time. Finally, a dynamic test was conducted on a road. The experiment showed that the detection speed of YOLOv4-Tiny in the dynamic test was higher than 60 FPS, and the accuracy rate surpassed 70%. Furthermore, the distance detection error of the 3D LiDAR was less than 3 cm, which is sufficiently accurate to be applied to complex environments on roads.