Limited Memory Resources Research Articles

FPGA–accelerated CNN for real-time plant disease identification

Using convolutional neural network (CNN) to identify plant diseases in-situ is a hot research topic in smart agriculture. Due to the memory-intensive and compute-intensive characteristics of CNN algorithm, it is difficult to implement CNN on edge terminals with limited memory and computational resources. In this paper, Field Programmable Gate Array (FPGA) is used to accelerate CNN to identify plant diseases. First, a 7-layer simple-structured network called “LiteCNN”, with only 176 K parameters and 78.47 M floating point operations (FLOPs) was designed. And knowledge distillation method was used to train LiteCNN, making that the accuracy reaches 95.24 %. Secondly, the acceleration circuit of LiteCNN was designed and implemented on “ZYNQ Z7-Lite 7020″ FPGA board. To compress the network and speed up plant disease identification, the following methods were applied: 1) Separable convolution took place of regular convolution, and a low-redundancy block convolution approach was used to load data; 2) The Batch Normalization (BN) layer was fused into the previous convolutional layer (or fully-connected layer); 3) Feature data and model parameters were expressed by half float data. As the basic function of the circuit achieved, methods including unrolling the for-loop, pipelining the for-loop, loop flattening and array partitioning were used to optimize the parallelism of the circuit. Finally, LiteCNN on the FPGA board was verified. The plant disease identification accuracy was 95.71 %, the inference speed was 0.071 s per frame, and the power consumption was 2.41 W. The results show that this paper proposed a low-power, high-accuracy and fast-speed plant disease identification terminal, which can be well applied for real-time plant disease identification in the field.

ILETC: Incremental learning for encrypted traffic classification using generative replay and exemplar

With the constant updating of applications and the emergence of various encryption technologies, it is important to achieve continuous learning of encrypted traffic. Traditional encrypted traffic classification techniques can only handle a fixed number of traffic classes and require all traffic data to be available at the same time, which consume a huge amount of memory to save the traffic data. In this paper, we propose an incremental learning method for encrypted traffic classification, called ILETC, to achieve continuous learning of encrypted traffic. ILETC can learn new encrypted traffic classes while avoiding forgetting knowledge of old encrypted traffic, with limited memory resources. It uses WGAN-GP to model the data distribution of encrypted traffic and design an exemplar set to select and store representative real traffic data. When learning from a new class of encrypted traffic, ILETC replays the knowledge of the old classes with the generated samples and exemplars to mitigate catastrophic forgetting. The ISCX VPN-nonVPN dataset and self-collected dataset are used to test the performance of ILETC. The results show that ILETC is superior to the state-of-the-art methods with an accuracy of over 98% on the ISCX dataset and over 94% in the two datasets together.

Tiny Machine Learning for High Accuracy Product Quality Inspection

The quality inspection of industrial products is a fundamental step in large-scale production as it boosts the yield and reduces the costs. Intelligent embedded platforms with built-in tiny machine learning (tinyML) algorithms and cameras can automate quality inspection; however, running complex deep learning algorithms in low-cost and low-power embedded devices is still challenging because of limited memory and energy resources. This article presents an innovative sensor system with three microcontroller unit (MCU)-based tinyML cameras capable of automatic artifact and anomaly detection in plastic components. The system consists of a top camera responsible for identifying shape defects and two side cameras for color anomalies. Data processing is executed locally with tinyML reducing data transmission to a few bytes. Two state-of-the-art convolutional neural network (CNN) architectures are evaluated, namely, MobileNetV2 and SqueezeNet. Results show how both the architectures—with appropriate compression techniques—are suitable to be evaluated by resource-constrained microcontrollers. The networks achieve 99% classification accuracy while maintaining suitable real-time performance, respectively, equal to 5 and 2 frames/s.

Energy-aware resource management in Internet of vehicles using machine learning algorithms

Internet of Vehicles (IoV) presents a new generation of vehicular communications with limited computation offloading, energy and memory resources with 5G/6G technologies that have grown enormously and are being used in wide variety of Intelligent Transportation Systems (ITS). Due to the limited battery power in smart vehicles, the concept of energy consumption is one of the main and critical challenges of the IoV environments. Optimizing resource management strategies for improving the energy consumption using AI-based methods is one of important solutions in the IoV environments. There are various machine learning algorithms for selecting optimal solutions for energy-efficient resource management strategies. This paper presents the existing energy-aware resource management strategies for the IoV case studies, and performs a comparative analysis among their applied AI-based methods and machine learning algorithms. This analysis presents a technical and deeper understanding of the technical aspects of existing machine learning and AI-based algorithms that will be helpful in design of new hybrid AI approaches for optimizing resource management strategies with reducing their energy consumption.

CoRAE: Energy Compaction-Based Correlation Pattern Recognition Training Using AutoEncoder

Automatic Target Recognition (ATR) using Correlation Pattern Recognition (CPR) in IoT-based applications encounters limitations like limited memory and inadequate computational resources. One reason is the required quantity of reference templates for each target/object to cover all features of a target/object. To mitigate the issue of reference templates per target/object without accuracy degradation, this paper proposes <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">energy compaction-based CPR autoencoder-training</i> . Additionally, a newly proposed performance metric known as <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Peak Energy Gain (PEG)</i> estimates the quality of the correlation plane and the feature compression capability CPR methods. The proposed, composite filtering strategy, Eigen Maximum Average Correlation Height (EMACH), and Extended Eigen Maximum Average Correlation Height ( <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">E</i> <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> MACH) are vigorously validated using publicly available biometric and object image databases. By training a single reference template, the proposed training method achieves 97.97% mean accuracy with the second-best approach of <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">E</i> <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> MACH that attains 53.04% mean accuracy on the Pose Estimation Database. For bio-metric fingerprint verification, the mean Equal Error Rate (EER) of the proposed approach and the composite strategy is 3% and 29.69%, respectively on the FVC2002DB1A database. Similarly, the mean EER of the proposed approach and the composite strategy is 10.55% and 26.32%, respectively on the FVC2006IA database. For FEI faces dataset, the proposed method achieves 1.41% mean EER, and the composite filtering approach achieves 21.43% mean EER. On the University of Tehran Iris database, the proposed autoencoder-based methodology obtains 19.07%, and 18.07% mean EER on the left and right side iris instances, respectively. The comparative results for each dataset demonstrate superiority of AE-based method over the state-of-the-art CPR methods.

Secure and efficient data exchange in Internet of Things networks – technology review

IoT devices are often accompanied by numerous hardware constraints, such as limited computing power, limited memory resources, and low network bandwidth. All these components have a significant impact on the difficulties in ensuring the security of IoT networks. For this reason, various solutions designed specifically for IoT devices have been developed to increase the level of security of such devices. The paper reviews the most popular solutions used in IoT networks that minimize problems caused by numerous limitations.

Data-Independent Structured Pruning of Neural Networks via Coresets.

Model compression is crucial for the deployment of neural networks on devices with limited computational and memory resources. Many different methods show comparable accuracy of the compressed model and similar compression rates. However, the majority of the compression methods are based on heuristics and offer no worst case guarantees on the tradeoff between the compression rate and the approximation error for an arbitrarily new sample. We propose the first efficient structured pruning algorithm with a provable tradeoff between its compression rate and the approximation error for any future test sample. Our method is based on the coreset framework, and it approximates the output of a layer of neurons/filters by a coreset of neurons/filters in the previous layer and discards the rest. We apply this framework in a layer-by-layer fashion from the bottom to the top. Unlike previous works, our coreset is data-independent, meaning that it provably guarantees the accuracy of the function for any input [Formula: see text], including an adversarial one.

Multi-task learning on the edge for effective gender, age, ethnicity and emotion recognition

More and more real-world applications, especially in cognitive robotics, require the running of different computer vision algorithms in parallel on board of embedded devices with limited GPU and memory resources. Multi-task learning, namely the usage of the same model to perform multiple classification and/or regression tasks by learning a shared low level representation, revealed to be a valid solution to reduce computation and required memory space while preserving the accuracy. In this paper, we propose a solution for real-time user profiling based on a multi-task convolutional neural network (CNN) for gender, age, ethnicity and emotion recognition from face images. To find the best trade-off between accuracy and processing time, we evaluate three different architectures, specifically designed for the purpose, and backbones, based on MobileNet, ResNet and SENet, which include convolutional layers, residual blocks and attention modules that already demonstrated great potential in face analysis. We trained the multi-task neural network with a custom learning procedure, which solves the problems of missing labels, dataset imbalance and loss function imbalance through label masking, batch balancing and a custom weighted loss function; there are no other multi-task neural networks for face analysis that address all these challenges simultaneously.. The proposed solution demonstrated its effectiveness in the comparison with the corresponding single-task CNNs in terms of accuracy, processing time and memory space; in fact, the multi-task CNNs achieved a processing speed-up between 2.5 and 4 times and a reduction of the memory space between 2 and 4 times, while preserving the accuracy. Moreover, the useful insights that arise from the experiments allow to choose a solution for face analysis easily integrable into real applications on smart cameras and embedded systems and most suited for the specific application constraints in terms of computational resources.

CAN3D: Fast 3D medical image segmentation via compact context aggregation.

Direct automatic segmentation of objects in 3D medical imaging, such as magnetic resonance (MR) imaging, is challenging as it often involves accurately identifying multiple individual structures with complex geometries within a large volume under investigation. Most deep learning approaches address these challenges by enhancing their learning capability through a substantial increase in trainable parameters within their models. An increased model complexity will incur high computational costs and large memory requirements unsuitable for real-time implementation on standard clinical workstations, as clinical imaging systems typically have low-end computer hardware with limited memory and CPU resources only. This paper presents a compact convolutional neural network (CAN3D) designed specifically for clinical workstations and allows the segmentation of large 3D Magnetic Resonance (MR) images in real-time. The proposed CAN3D has a shallow memory footprint to reduce the number of model parameters and computer memory required for state-of-the-art performance and maintain data integrity by directly processing large full-size 3D image input volumes with no patches required. The proposed architecture significantly reduces computational costs, especially for inference using the CPU. We also develop a novel loss function with extra shape constraints to improve segmentation accuracy for imbalanced classes in 3D MR images. Compared to state-of-the-art approaches (U-Net3D, improved U-Net3D and V-Net), CAN3D reduced the number of parameters up to two orders of magnitude and achieved much faster inference, up to 5 times when predicting with a standard commercial CPU (instead of GPU). For the open-access OAI-ZIB knee MR dataset, in comparison with manual segmentation, CAN3D achieved Dice coefficient values of (mean = 0.87±0.02 and 0.85±0.04) with mean surface distance errors (mean = 0.36±0.32mm and 0.29±0.10mm) for imbalanced classes such as (femoral and tibial) cartilage volumes respectively when training volume-wise under only 12G video memory. Similarly, CAN3D demonstrated high accuracy and efficiency on a pelvis 3D MR imaging dataset for prostate cancer consisting of 211 examinations with expert manual semantic labels (bladder, body, bone, rectum, prostate) now released publicly for scientific use as part of this work.

Robust and Efficient Star Identification Algorithm based on 1-D Convolutional Neural Network

As the core of the attitude determination system, the star sensor working in “lost in space” scenarios requires the star identification algorithm to be robust and fast with limited computing and memory resources. Nevertheless, previous algorithms are not satisfactory in robustness and identification speed. Hence, motivated by the fact that the one-dimensional convolutional neural network (1D-CNN) is suitable for sequential data, this article proposes a robust and efficient star identification algorithm, where 1D-CNN is used to process mixed initial features from star points. Moreover, this article proposes a combined star points selection strategy technique and a mixed initial features extraction technique to further improve the performance of 1D-CNN-based algorithm. Experimental results show that, compared with the state-of-the-art algorithm, the proposed algorithm can improve the average identification accuracy by 0.76%, the identification speed by 1.86× with the comparable memory consumption.

Improved Knowledge Distillationwith Dynamic Network Pruning

Deploying convolutional neural networks to mobile or embedded devices is often prohibited by limited memory and computational resources. This is particularly problematic for the most successful networks, which tend to be very large and require long inference times. Many alternative approaches have been developed for compressing neural networks based on pruning, regularization, quantization or distillation. In this paper, we propose the “Knowledge Distillation with Dynamic Pruning” (KDDP), which trains a dynamically pruned compact student network under the guidance of a large teacher network. In KDDP, we train the student network with supervision from the teacher network, while applying L1 regularization on the neuron activations in a fully-connected layer. Subsequently, we prune inactive neurons. Our method automatically determines the final size of the student model. We evaluate the compression rate and accuracy of the resulting networks on an image classification dataset, and compare them to results obtained by Knowledge Distillation (KD). Compared to KD, our method produces better accuracy and more compact models.

STAR-Lite: A light-weight scalable self-taught learning framework for older adults’ activity recognition

Recognizing activities for older adults is challenging as we observe a variety of activity patterns caused due to aging (e.g., limited dexterity, limb control, slower response time) or/and underlying health conditions (e.g., dementia). However, existing literature with deep learning methods has successfully recognized activities when the dataset contains high-quality annotations and is captured in a controlled environment. On the contrary, data captured in a real-world environment, especially with older adults exhibiting memory-related symptoms, varying psychological and mental health status, reliance on caregivers to perform daily activities, and unavailability of domain-specific annotators, makes obtaining quality data with annotations challenging; leaving us with limited labeled data and abundant unlabeled data. In this paper, we hypothesize that projecting the labeled data representations comprising a specific set of activities onto a new representation space characterized by the unlabeled data comprising activities beyond the limited activities in the labeled dataset would help us rely less on the annotated data to improve activity detection performance. Motivated by this, we propose STAR-Lite, a self-taught learning framework that involves a pre-training framework to prepare the new representation space considering activities beyond the initial labels in the labeled dataset. STAR-Lite projects the labeled data representations on the new representation space characterized by unlabeled data labels and learns higher-level representations of the labeled dataset while optimizing inter- and intra- class distances without explicitly using a computation hungry similarity-based approach. We demonstrate that our proposed approach, STAR-Lite (a) improves activity recognition performance in a supervised setting and (b) is feasible for real-world deployment. To enhance the feasibility of deploying STAR-Lite on devices with limited memory resources, we explore model compression techniques such as pruning and quantization and propose a novel layer-wise pruning-rate optimization technique that effectively compresses the network while preserving the model performance. The evaluation was performed using the Alzheimer’s Activity Recognition dataset (AAR) captured from 25 individuals living in a retirement community center with IRB approval (#Y18NR12035) using an in-house SenseBox infrastructure while concurrently assessing the clinical evaluation of the participants for dementia, and independent living. Our extensive evaluation reveals that STAR-Lite can detect activities with an F1-score of 85.12% despite 62% reduction in model size and 5% improvement of execution time on a resource constrained device.

Combining multi-view ensemble and surrogate lagrangian relaxation for real-time 3D biomedical image segmentation on the edge

Real-time 3D biomedical image segmentation is always preferred considering the exponentially growing medical imaging data for the past decade. Recently deep learning has significantly boosted the performance of automatic medical image segmentation with high computation and memory requirements, especially for 3D biomedical images. Meanwhile, the privacy and security of patient data have always been the primary concern in medical applications among hospitals and clinics, and there also exists some applications which need real-time processing in clinic practice. Thus, 3D biomedical image segmentation is typically required to be performed locally (i.e. on the edge) with limited computation and memory resources. In this paper, we propose to combine multi-view ensemble and Surrogate Lagrangian relaxation (SLR) for real-time 3D biomedical image segmentation on the edge. Instead of directly dealing with 3D biomedical images, our segmentation conducts on the three 2D domains of the 3D images with an ensemble strategy. In addition, Surrogate Lagrangian relaxation is proposed to compress the model to enable high efficiency and real-time processing. Experiments on a typical edge Nvidia GPU show that our method achieves real-time processing which is 1.5× faster with an improvement of 9% on accuracy compared with single-view models. It also saves 26× computational resources and 6× memory resources compared to 3D segmentation models.

TinyIREE: An ML Execution Environment for Embedded Systems From Compilation to Deployment

Machine learning model deployment for training and execution has been an important topic for industry and academic research in the last decade. Much of the attention has been focused on developing specific toolchains to support acceleration hardware. In this article, we present Intermediate Representation Execution Environment (IREE), a unified compiler and runtime stack with the explicit goal to scale down machine learning programs to the smallest footprints for mobile and edge devices, while maintaining the ability to scale up to larger deployment targets. IREE adopts a compiler-based approach and optimizes for heterogeneous hardware accelerators through the use of the Multi-Level IR (MLIR) compiler infrastructure, which provides the means to quickly design and implement multilevel compiler intermediate representations (IR). More specifically, this article is focused on TinyIREE, which is a set of deployment options in IREE that accommodate the limited memory and computation resources in embedded systems and bare-metal platforms, while also demonstrating IREE’s intuitive workflow that generates workloads for different ISA extensions and application binary interface (ABIs) through LLVM.

Automatic Distributed Deep Learning Using Resource-Constrained Edge Devices

Processing data generated at high volume and speed from the Internet of Things, smart cities, domotic, intelligent surveillance, and e-healthcare systems require efficient data processing and analytics services at the Edge to reduce the latency and response time of the applications. The fog computing edge infrastructure consists of devices with limited computing, memory, and bandwidth resources, which challenge the construction of predictive analytics solutions that require resource-intensive tasks for training machine learning models. In this work, we focus on the development of predictive analytics for urban traffic. Our solution is based on deep learning techniques localized in the Edge, where computing devices have very limited computational resources. We present an innovative method for efficiently training the gated recurrent-units (GRUs) across available resource-constrained CPU and GPU Edge devices. Our solution employs distributed GRU model learning and dynamically stops the training process to utilize the low-power and resource-constrained Edge devices while ensuring good estimation accuracy effectively. The proposed solution was extensively evaluated using low-powered ARM-based devices, including Raspberry Pi v3 and the low-powered GPU-enabled device NVIDIA Jetson Nano, and also compared them with Single-CPU Intel Xeon machines. For the evaluation experiments, we used real-world Floating Car Data. The experiments show that the proposed solution delivers excellent prediction accuracy and computational performance on the Edge when compared to the baseline methods.

On-sensor binarized CNN inference with dynamic model swapping in pixel processor arrays.

Many types of Convolutional Neural Network (CNN) models and training methods have been proposed in recent years aiming to provide efficiency for embedded and edge devices with limited computation and memory resources. The wide variety of architectures makes this a complex task that has to balance generality with efficiency. Among the most interesting camera-sensor architectures are Pixel Processor Arrays (PPAs). This study presents two methods that are useful for embedded CNNs in general but particularly suitable for PPAs. The first is for training purely binarized CNNs, the second is for deploying larger models with a model swapping paradigm that loads model components dynamically. Specifically, this study trains and implements networks with batch normalization and adaptive threshold for binary activations. Then, we convert batch normalization and binary activations into a bias matrix which can be parallelly implemented by an add/sub operation. For dynamic model swapping, we propose to decompose applications that are beyond the capacity of a PPA into sub-tasks that can be solved by tree networks that can be loaded dynamically as needed. We demonstrate our approaches to various tasks including classification, localization, and coarse segmentation on a highly resource constrained PPA sensor-processor.

Recognition of Aircraft Activities at Airports on Video Micro-Satellites: Methodology and Experimental Validation

The remote sensing satellite constellation based on micro-satellites is an important means to construct a global and all-sky earth observation system in the future. Therefore, realizing the recognition of aircraft activities on video micro-satellites is a key technology that needs to be solved urgently. In this paper, an efficient algorithm for aircraft activity recognition that can be deployed on video micro-satellites was proposed. First, aircraft detection was performed on the first incoming remote sensing image using a robust DCNN-based object detection model. Then, a multi-target tracking model incorporating geospatial information was built for aircraft tracking and activity recognition. The algorithm was deployed on an embedded AI computer which was a COTS component. The algorithm was verified using remote sensing videos from commercial micro-satellites. Experimental results show that the algorithm can process aircraft targets of different sizes, and is equally effective even with complex environmental backgrounds, lighting conditions, and various movements of the aircraft, such as turning, entering, and exiting. Based on aircraft tracking results and geospatial information, the motion speed of each aircraft can be obtained, and its activity can be divided into parking, taxiing, or flying. The scheme proposed in this paper has good application prospects in the realization of on-orbit event recognition in micro-satellites with limited computing and memory resources.

YOLOv5-R: lightweight real-time detection based on improved YOLOv5

In reality, deploying the traditional object detection algorithm on mobile and embedded devices is difficult due to the limited memory and computation resources. To solve this problem, we propose a lightweight, real-time detection algorithm, YOLOv5-R. First, inserting the efficient channel attention (ECA) module into the Ghost module achieves the information interaction between channels and suppresses the redundant features. And applying dense connections to it further improves feature reuse and network performance, forming a Ghost-ECA-Dense (GED) module. Utilizing the GED module to construct the F-GED, which is the backbone with fewer parameters and better performance, replace CSPDarknet53 in YOLOv5. Second, the redundant operations are replaced with Ghost modules and depthwise separable convolution in the neck and head of YOLOv5, and YOLOv5-R is constructed, which significantly reduces the network size and improves the inference speed. Finally, YOLOv5-R is deployed into the AI embedded device Jetson Nano for TensorRT acceleration. The experimental results indicated that the performance of YOLOv5-R outperformed the YOLOv5s. Specifically, the mean average precision was increased from 69.9% to 72.7%, the number of floating-point operations per second and parameters are reduced by 25% and 47.2%, respectively. The model weight was only 7.52 MB, and frames per second achieved 37, which can provide real-time detection.

Lightweight Hardware Architecture for Object Detection in Driver Assistance Systems

Object detection on hardware platforms plays a very significant role in developing driver assistance systems (DASs) with limited computational resources. Object detection for DAS is a multiclass detection problem that involves detecting various objects like cars, auto, traffic lights, bicycles, pedestrians, etc. DAS also requires accuracy, speed, and sensitivity for detecting these objects in various challenging conditions. The lighting and weather conditions pose a serious challenge for accurate object detection for DAS. This paper proposes a speed-efficient and lightweight fully convolutional neural network (CNN) architecture for object detection in adverse rainy conditions. The proposed architecture uses a CNN-based deraining network with a custom SSIM loss function in the object detection pipeline, which can give an accurate performance using limited computational and memory resources. The object detection architecture contains some architectural modifications to the existing single shot multibox detector (SSD) architecture to make it more hardware efficient and improve accuracy on small objects. It uses a trainable color transformation module using [Formula: see text] convolutions for handling the adverse lighting conditions encountered in DAS. The architecture uses feature fusion and the dilated convolution approach to enhance the accuracy of the proposed architecture on small objects. The datasets available for object detection in DAS are very imbalanced with cars as a predominant object. The class weight penalization technique is used to improve the performance of the architecture on scarcely present objects. The performance of the architecture is evaluated on well-known datasets like Kitti, Udacity, Indian Driving Dataset (IDD), and DAWN. The architecture achieves satisfactory performance in terms of mean average precision (mAP) and detection time on all these datasets. It requires three times fewer hardware resources compared to existing architectures. The lightweight nature of the proposed architecture and modification of CNN architecture with TensorRT allow the efficient implementation on the jetson nanohardware platform for prototyping, which can be integrated with other intelligent transportation systems.

GhostNets on Heterogeneous Devices via Cheap Operations

Deploying convolutional neural networks (CNNs) on mobile devices is difficult due to the limited memory and computation resources. We aim to design efficient neural networks for heterogeneous devices including CPU and GPU, by exploiting the redundancy in feature maps, which has rarely been investigated in neural architecture design. For CPU-like devices, we propose a novel CPU-efficient Ghost (C-Ghost) module to generate more feature maps from cheap operations. Based on a set of intrinsic feature maps, we apply a series of linear transformations with cheap cost to generate many ghost feature maps that could fully reveal information underlying intrinsic features. The proposed C-Ghost module can be taken as a plug-and-play component to upgrade existing convolutional neural networks. C-Ghost bottlenecks are designed to stack C-Ghost modules, and then the lightweight C-GhostNet can be easily established. We further consider the efficient networks for GPU devices. Without involving too many GPU-inefficient operations (e.g.,, depth-wise convolution) in a building stage, we propose to utilize the stage-wise feature redundancy to formulate GPU-efficient Ghost (G-Ghost) stage structure. The features in a stage are split into two parts where the first part is processed using the original block with fewer output channels for generating intrinsic features, and the other are generated using cheap operations by exploiting stage-wise redundancy. Experiments conducted on benchmarks demonstrate the effectiveness of the proposed C-Ghost module and the G-Ghost stage. C-GhostNet and G-GhostNet can achieve the optimal trade-off of accuracy and latency for CPU and GPU, respectively. Code is available at https://github.com/huawei-noah/CV-Backbones.