Faster Convolutional Neural Network Research Articles

Fast Mask Recurrent Convolutional Neural Network for IoT-Based Maternal and Fetal Monitoring in High-Risk Pregnancies

Fast Mask Recurrent Convolutional Neural Network for IoT-Based Maternal and Fetal Monitoring in High-Risk Pregnancies

Improving Badminton Player Detection Using YOLOv3 with Different Training Heuristic

There has been a considerable rise in the amount of research and development focused on computer vision over the previous two decades. One of the most critical processes in computer vision is "visual tracking," which involves following objects with a camera. Tracking objects is the practice of following an individual moving object or group of moving things over time. Identifying or connecting target elements in consecutive video frames of a badminton match requires visual object tracking. The aim of this study is to identify badminton players using the You Only Look Once (YOLO) technique in conjunction with a variety of training heuristics. This methodology has a few advantages over other approaches to detecting objects. The convolutional neural network and Fast convolutional neural network are two examples of the many algorithmic approaches that are available. In this study, a neural network is used to produce predictions about the bounding boxes and the class probabilities for these boxes.. The results demonstrated that it was far faster than other methods in terms of its ability to recognize the image. The performance of image classification networks significantly improved as a result of the implementation of a variety of training strategies for the detection of objects. The mean average precision score for YOLOv3 with various training heuristics increased from 32.0 to 36.0 as a direct result of these adjustments. In comparison to YOLOv3, our future study might examine the performance of alternative models like Faster R-CNN or RetinaNet.

Aggregate Channel Features and Fast Regions CNN Approach for Classification of Ship and Iceberg

Detection and classification of icebergs and ships in synthetic aperture radar (SAR) images play a vital role in marine surveillance systems even though available adaptive threshold methods give satisfying results on detection and classification for ships and icebergs, including techniques of convolutional neural networks (CNNs), but need more accuracy and precision. An efficient and accurate method was developed to detect and classify the ship and icebergs. Hence, the research method proposed locating and classifying both ships and icebergs in a given SAR image with the help of deep learning (DL) and non-DL methods. A non-DL method utilized here was the aggregate channel features (ACF) detector, which extracts region proposals from huge SAR images. The DL object detector called fast regions CNN (FRCNN) detects objects accurately from the result of ACF since the ACF method avoids unwanted regions. The novelty of this study was that ACF-FRCNN concentrates only on accurately classifying ships and icebergs. The proposed ACF-FRCNN method gave a better performance in terms of loss (18.32%), accuracy (96.34%), recall (98.32%), precision (95.97%), and the F1 score (97.13%). Compared to other conventional methods, the combined effect of ACF and FRCNN increased the speed and quality of the detection of ships and icebergs. Thus, the ACF-FRCNN method is considered a novel method for over 75 × 75 resolution ship and iceberg SAR images.

Deep Learning-Based Classification and Feature Extraction for Predicting Pathogenesis of Foot Ulcers in Patients with Diabetes.

The World Health Organization (WHO) has identified that diabetes mellitus (DM) is one of the most prevalent disease worldwide. Individuals with DM have a higher risk of mortality, and it is crucial to prioritize the treatment of foot ulcers, which is a significant complication associated with the disease, as they lead to the development of plantar ulcers, which results in the need to amputate part of the foot or leg. People with diabetes are at risk of experiencing various complications, such as heart disease, eye problems, kidney dysfunction, nerve damage, skin issues, foot ulcers, and dental diseases. Unawareness of the risk associated with diabetic foot ulcers (DFU) is a significant contributing factor to the mortality of diabetic patients. Evolving technological advancements such as deep learning techniques can be used to predict the symptoms of diabetic foot ulcers as early as possible, which helps to provide effective treatment to DM patients. This research introduces a methodology for analyzing images of foot ulcers in diabetic patients, focusing on feature extraction and classification. The dataset used in this study was collected from historical medical records and foot images of patients with diabetes, who commonly experience foot ulcers as a major complication. The dataset was pre-processed and segmented, and features were extracted using a deep recurrent neural network (DRNN). Image and numerical/text data were extracted separately, and the normal and abnormal diabetes ranges were identified. Foot images of patients with abnormal diabetes ranges were separated and classified using a pre-trained fast convolutional neural network (PFCNN) with U++net. The classification procedure involves the analysis of foot ulcers to predict their pathogenesis. To assess the effectiveness of the proposed technique, the study presented simulation results, including a confusion matrix and receiver operating characteristic curve. These results specifically focused on predicting two classes: normal and abnormal diabetes foot ulcerations. The analysis yielded various parameters, including accuracy, precision, recall curve, and area under the curve. The main goal of the study was to introduce an novel technique for assessing the risk of foot ulceration development in patients with diabetes, leveraging the analysis of foot ulcer images. The researchers collected a dataset of foot images and medical data from historical records of patients with diabetes and pre-processed and segmented the data. They then used a deep recurrent neural network to extract features from the segmented data and identified normal and abnormal diabetes ranges based on numerical and text data. Foot images of patients with abnormal diabetes ranges were classified using a pre-trained fast convolutional neural network with U++net to examine foot ulcers and forecast the development of the risk of diabetic foot ulcers (DFU). The study assessed the accuracy of the proposed technique as 99.32% by simulating results for feature extraction and the classification of diabetic foot ulcers. A comparison was made between this proposed technique and existing approaches.

Transient velocity field prediction and uncertainty quantification of pump-jet propulsor using variational Bayesian neural networks

The study provides a fast and accurate variational Bayesian convolution neural network (VB–CNN) framework to obtain the pump-jet propulsor (PJP) transient velocity field. In practical engineering or experiments, it can obtain the transient velocity field of PJP only by discrete pressure points for different times conditions. Compared with the computational fluid dynamics (CFD) method, it can save a lot of computing time and computing resources. This study also uses the VB-CNN method to solve the problem of poor performance of CNN in small data-sets, VB-CNN method can consider the prior knowledge of PJP transient data to improve the prediction accuracy. The predicted values, errors and variances of the PJP transient wake velocity field and stator domain at different times and different rotating speed conditions are investigated. The uncertainty of the transient flow field is analyzed using 95% confidence intervals. It also uses GPU technology to improve training speed. The research results show that VB-CNN and CNN methods can better predict the PJP transient flow field, while the VB-CNN method has higher accuracy than CNN method. The velocity gradient and velocity profile of the PJP transient velocity field are predicted accurately, the predicted velocity field is in good agreement with the CFD flow field. All linear correlation coefficients exceeds 0.99. The value of structural similarity index (SSIM) at different times exceeds 0.95.

A novel adaptive network intrusion detection system for internet of things.

Cyber-attack is one of the most challenging aspects of information technology. After the emergence of the Internet of Things, which is a vast network of sensors, technology started moving towards the Internet of Things (IoT), many IoT based devices interplay in most of the application wings like defence, healthcare, home automation etc., As the technology escalates, it gives an open platform for raiders to hack the network devices. Even though many traditional methods and Machine Learning algorithms are designed hot, still it "Have a Screw Loose" in detecting the cyber-attacks. To "Pull the Plug on" an effective "Intrusion Detection System (IDS)" is designed with "Deep Learning" technique. This research work elucidates the importance in detecting the cyber-attacks as "Anomaly" and "Normal". Fast Region-Based Convolution Neural Network (Fast R-CNN), a deep convolution network is implemented to develop an efficient and adaptable IDS. After hunting many research papers and articles, "Gradient Boosting" is found to be a powerful optimizer algorithm that gives us a best results when compared to other existing methods. This algorithm uses "Regression" tactics, a statistical technique to predict the continuous target variable that correlates between the variables. To create a structured valid dataset, a stacked model is made by implementing the two most popular dimensionality reduction techniques Principal Component Analysis (PCA) and Singular Value Decomposition (SVD) algorithms. The brainwaves made us to hybridize Fast R-CNN and Gradient Boost Regression (GBR) which reduces the loss function, processing time and boosts the model's performance. All the above said methods are trained and tested with NIDS dataset V.10 2017. Finally, the "Decision Making" model decides the best result by giving an alert message. Our proposed model attains a high accuracy of 99.5% in detecting the "Cyber Attacks". The experiment results revealed that the effectiveness of our proposed model surpasses other deep neural network and machine learning techniques which have less accuracy.

Real Time Road Traffic Detection Using Computer Vision

Object detection using Machine learning has achieved very good performance but there are many problems with images in real-world shooting such as noise, blurring or rotating jitter, etc. These problems effects the object detection. The main objective is to detect objects using You Only Look Once (YOLO) algorithm. The YOLO method has several advantages when compared to other object detection algorithms. In other algorithms like Convolutional Neural Network (CNN), Fast-Convolutional Neural Network the algorithm will not look at the image completely, but in YOLO, the algorithm looks the image completely by predicting the bounding boxes using convolutional network and finds class probabilities for these boxes and also detects the image faster as compared to other algorithms. We have used this algorithm for detecting different types of objects.

Design of a Low-Power Super-Resolution Architecture for Virtual Reality Wearable Devices

Head-mounted displays (HMDs) have made a virtual reality (VR) accessible to a widespread consumer market, introducing a revolution in many applications. Among the limitations of current HMD technology, the need for generating high-resolution images and streaming them at adequate frame rates is one of the most critical. Super-resolution (SR) convolutional neural networks (CNNs) can be exploited to alleviate timing and bandwidth bottlenecks of video streaming by reconstructing high-resolution images locally (i.e., near the display). However, such techniques involve a significant amount of computations that makes their deployment within area-/power-constrained wearable devices often unfeasible. This research work originated from the consideration that the human eye can capture details with high acuity only within a certain region, called the fovea. Therefore, we designed a custom hardware architecture able to reconstruct high-resolution images by treating foveal region (FR) and peripheral region (PR) through accurate and inaccurate operations, respectively. Hardware experiments demonstrate the effectiveness of our proposal: a customized fast SR CNN (FSRCNN) accelerator realized as described here and implemented on a 28-nm process technology is able to process up to 214 ultrahigh definition frames/s, while consuming just 0.51 pJ/pixel without compromising the perceptual visual quality, thus achieving a 55% energy reduction and a <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\times 14$ </tex-math></inline-formula> times higher throughput rate, with respect to state-of-the-art competitors.

Deep-learning convolutional neural network-based scatter correction for contrast enhanced digital breast tomosynthesis in both cranio-caudal and mediolateral-oblique views.

Scatter radiation in contrast-enhanced digital breast tomosynthesis (CEDBT) reduces the image quality and iodinated lesion contrast. Monte Carlo simulation can provide accurate scatter estimation at the cost of computational burden. A model-based convolutional method trades off accuracy for processing speed. The purpose of this study is to develop a fast and robust deep-learning (DL) convolutional neural network (CNN)-based scatter correction method for CEDBT. Projection images and scatter maps of digital anthropomorphic breast phantoms were generated using Monte Carlo simulations. Experiments were conducted to validate the simulated scatter-to-primary ratio (SPR) at different locations of a breast phantom. Simulated projection images were used for CNN training and testing. Two separate U-Nets [low-energy (LE)-CNN and high-energy (HE)-CNN] were trained for LE and HE spectrum, respectively. CNN-based scatter correction was applied to a clinical case with a malignant iodinated mass to evaluate the influence on the lesion detection. The average and standard deviation of mean absolute percentage error of LE-CNN and HE-CNN estimated scatter map are and , respectively. For clinical cases, the lesion signal difference to noise ratio average improvement was 190% after CNN-based scatter correction. To conduct scatter correction on clinical CEDBT images, the whole process of loading CNNs parameters and scatter correction for LE and HE images took , with 9GB GPU computational cost. The SPR variation across the breast agrees between experimental measurements and Monte Carlo simulations. We developed a CNN-based scatter correction method for CEDBT in both CC view and mediolateral-oblique view with high accuracy and fast speed.

Breast Cancer Detection Using Image Processing and Machine Learning

Different breast cancer detection systems have been developed to help clinicians analyze screening mammograms. Breast cancer has been increasing gradually so scientists work to develop new methods to reduce the risks of this life-threatening disease. Convolutional Neural Networks (CNNs) have shown much promise In the field of medical imaging because of recent developments in deep learning. However, CNN’s based methods have been restricted due to the small size of the few public breast cancer datasets. This research has developed a new framework and introduced it to detect breast cancer. This framework utilizes Convolutional Neural Networks (CNNs) and image processing to achieve its goal because CNNs have been an important success in image recognition, reaching human performance. An efficient and fast CNN pre-trained object detector called RetinaNet has been used in this research. RetinaNet is an uncomplicated one-stage object detector. A two-stage transfer learning has been used with the selected detector to improve the performance. RetinaNet model is initially trained with a general-purpose dataset called COCO dataset. The transfer learning is then used to apply the RetinaNet model to another dataset of mammograms called the CBIS-DDSM dataset. Finally, the second transfer learning is used to test the RetinaNet model onto a small dataset of mammograms called the INbreast dataset. The results of the proposed two-stage transfer learning (RetinaNet → CBIS-DDSM → INbreast) are better than the other state-of-the-art methods on the public INbreast dataset. Furthermore, the True Positive Rate (TPR) is 0.99 ± 0.02 at 1.67 False Positives per Image (FPPI), which is better than the one-stage transfer learning with a TPR of 0.94 ± 0.02 at 1.67 FPPI.

Fetal electrocardiogram signal extraction and analysis method combining fast independent component analysis algorithm and convolutional neural network

Fetal electrocardiogram (ECG) signals provide important clinical information for early diagnosis and intervention of fetal abnormalities. In this paper, we propose a new method for fetal ECG signal extraction and analysis. Firstly, an improved fast independent component analysis method and singular value decomposition algorithm are combined to extract high-quality fetal ECG signals and solve the waveform missing problem. Secondly, a novel convolutional neural network model is applied to identify the QRS complex waves of fetal ECG signals and effectively solve the waveform overlap problem. Finally, high quality extraction of fetal ECG signals and intelligent recognition of fetal QRS complex waves are achieved. The method proposed in this paper was validated with the data from the PhysioNet computing in cardiology challenge 2013 database of the Complex Physiological Signals Research Resource Network. The results show that the average sensitivity and positive prediction values of the extraction algorithm are 98.21% and 99.52%, respectively, and the average sensitivity and positive prediction values of the QRS complex waves recognition algorithm are 94.14% and 95.80%, respectively, which are better than those of other research results. In conclusion, the algorithm and model proposed in this paper have some practical significance and may provide a theoretical basis for clinical medical decision making in the future.

Polarization Orientation Method Based on Remote Sensing Image in Cloudy Weather

Autonomous navigation technology is a core technology for intelligent operation, allowing the vehicles to perform tasks without relying on external information, which effectively improves the concealability and reliability. In this paper, based on the previous research on the bionic compound eye, a multi-channel camera array with different polarization degrees was used to construct the atmospheric polarization state measurement platform. A polarization trough threshold segmentation algorithm was applied to study the distribution characteristics and characterization methods of polarization states in atmospheric remote sensing images. In the extracted polarization feature map, the tilting suggestion box was obtained based on the multi-direction window extraction network (similarity-based region proposal networks, SRPN) and the rotation of the suggestion box (Rotation Region of interests, RRoIs). Fast Region Convolutional Neural Networks (RCNN) was used to screen the suggestion boxes, and the Non-maximum suppression (NMS) method was used to select the angle, corresponding to the label of the suggestion box with the highest score, as the solar meridian azimuth in the vehicle coordinate system. The azimuth angle of the solar meridian in the atmospheric coordinate system can be calculated by the astronomical formula. Finally, the final heading angle can be obtained according to the conversion relationship between the coordinate systems. By fitting the measured data based on the least Square method, the slope K value is −1.062, RMSE (Root Mean Square Error) is 6.984, and the determination coefficient R-Square is 0.9968. Experimental results prove the effectiveness of the proposed algorithm, and this study can construct an autonomous navigation algorithm with high concealment and precision, providing a new research idea for the research of autonomous navigation technology.

Fast Image Super-Resolution Using Particle Swarm Optimization-Based Convolutional Neural Networks.

Image super-resolution based on convolutional neural networks (CNN) is a hot topic in image processing. However, image super-resolution faces significant challenges in practical applications. Improving its performance on lightweight architectures is important for real-time super-resolution. In this paper, a joint algorithm consisting of modified particle swarm optimization (SMCPSO) and fast super-resolution convolutional neural networks (FSRCNN) is proposed. In addition, a mutation mechanism for particle swarm optimization (PSO) was obtained. Specifically, the SMCPSO algorithm was introduced to optimize the weights and bias of the CNNs, and the aggregation degree of the particles was adjusted adaptively by a mutation mechanism to ensure the global searching ability of the particles and the diversity of the population. The results showed that SMCPSO-FSRCNN achieved the most significant improvement, being about 4.84% better than the FSRCNN model, using the BSD100 data set at a scale factor of 2. In addition, a chest X-ray super-resolution images classification test experiment was conducted, and the experimental results demonstrated that the reconstruction ability of this model could improve the classification accuracy by 13.46%; in particular, the precision and recall rate of COVID-19 were improved by 45.3% and 6.92%, respectively.

Analyzing students' attention by gaze tracking and object detection in classroom teaching

PurposeAttention is one of the most important factors to affect the academic performance of students. Effectively analyzing students' attention in class can promote teachers' precise teaching and students' personalized learning. To intelligently analyze the students' attention in classroom from the first-person perspective, this paper proposes a fusion model based on gaze tracking and object detection. In particular, the proposed attention analysis model does not depend on any smart equipment.Design/methodology/approachGiven a first-person view video of students' learning, the authors first estimate the gazing point by using the deep space–time neural network. Second, single shot multi-box detector and fast segmentation convolutional neural network are comparatively adopted to accurately detect the objects in the video. Third, they predict the gazing objects by combining the results of gazing point estimation and object detection. Finally, the personalized attention of students is analyzed based on the predicted gazing objects and the measurable eye movement criteria.FindingsA large number of experiments are carried out on a public database and a new dataset that is built in a real classroom. The experimental results show that the proposed model not only can accurately track the students' gazing trajectory and effectively analyze the fluctuation of attention of the individual student and all students but also provide a valuable reference to evaluate the process of learning of students.Originality/valueThe contributions of this paper can be summarized as follows. The analysis of students' attention plays an important role in improving teaching quality and student achievement. However, there is little research on how to automatically and intelligently analyze students' attention. To alleviate this problem, this paper focuses on analyzing students' attention by gaze tracking and object detection in classroom teaching, which is significant for practical application in the field of education. The authors proposed an effectively intelligent fusion model based on the deep neural network, which mainly includes the gazing point module and the object detection module, to analyze students' attention in classroom teaching instead of relying on any smart wearable device. They introduce the attention mechanism into the gazing point module to improve the performance of gazing point detection and perform some comparison experiments on the public dataset to prove that the gazing point module can achieve better performance. They associate the eye movement criteria with visual gaze to get quantifiable objective data for students' attention analysis, which can provide a valuable basis to evaluate the learning process of students, provide useful learning information of students for both parents and teachers and support the development of individualized teaching. They built a new database that contains the first-person view videos of 11 subjects in a real classroom and employ it to evaluate the effectiveness and feasibility of the proposed model.

Accelerating Deep Convolutional Neural Networks Using Number Theoretic Transform

Modern deep convolutional neural networks (CNNs) suffer from high computational complexity due to excessive convolution operations. Recently, fast convolution algorithms such as fast Fourier transform (FFT) and Winograd transform have gained attention to address this problem. They reduce the number of multiplications required in the convolution operation by replacing it with element-wise multiplication in the transform domain. However, fast convolution-based CNN accelerators have three major concerns: expensive domain transform, large memory overhead, and limited flexibility in kernel size. In this paper, we present a novel CNN accelerator based on number theoretic transform (NTT), which overcomes the existing limitations. We propose the low-cost NTT and inverse-NTT converter that only use adders and shifters for on-chip domain transform, which solves the inflated bandwidth problem and enables more parallel computations in the accelerator. We also propose the accelerator architecture that includes multiple tile engines with the optimized data flow and mapping. Finally, we implement the proposed NTT-based CNN accelerator on the Xilinx Alveo U50 FPGA and evaluate it for popular deep CNN models. As a result, the proposed accelerator achieves 2859.5, 990.3, and 805.6 GOPS throughput for VGG-16, GoogLeNet, and Darknet-19, respectively. It outperforms the existing fast convolution-based CNN accelerators up to <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$9.6\times $ </tex-math></inline-formula> .

Enhancement of classification and prediction accuracy for breast cancer detection using fast convolution neural network with ensemble algorithm

Enhancement of classification and prediction accuracy for breast cancer detection using fast convolution neural network with ensemble algorithm

Segmentation Based Real Time Anomaly Detection and Tracking Model for Pedestrian Walkways

Presently, video surveillance is commonly employed to ensure security in public places such as traffic signals, malls, railway stations, etc. A major challenge in video surveillance is the identification of anomalies that exist in it such as crimes, thefts, and so on. Besides, the anomaly detection in pedestrian walkways has gained significant attention among the computer vision communities to enhance pedestrian safety. The recent advances of Deep Learning (DL) models have received considerable attention in different processes such as object detection, image classification, etc. In this aspect, this article designs a new Panoptic Feature Pyramid Network based Anomaly Detection and Tracking (PFPN-ADT) model for pedestrian walkways. The proposed model majorly aims to the recognition and classification of different anomalies present in the pedestrian walkway like vehicles, skaters, etc. The proposed model involves panoptic segmentation model, called Panoptic Feature Pyramid Network (PFPN) is employed for the object recognition process. For object classification, Compact Bat Algorithm (CBA) with Stacked Auto Encoder (SAE) is applied for the classification of recognized objects. For ensuring the enhanced results better anomaly detection performance of the PFPN-ADT technique, a comparison study is made using University of California San Diego (UCSD) Anomaly data and other benchmark datasets (such as Cityscapes, ADE20K, COCO), and the outcomes are compared with the Mask Recurrent Convolutional Neural Network (RCNN) and Faster Convolutional Neural Network (CNN) models. The simulation outcome demonstrated the enhanced performance of the PFPN-ADT technique over the other methods.

Designing Very Fast and Accurate Convolutional Neural Networks With Application in ICD and Smart Electrocardiograph Devices

An implantable cardioverter-defibrillator (ICD) is a device that must detect VT and VF arrhythmias on time and treat them. In this project, three CNN networks are designed to introduce the practical methods of using deep learning in heart electrophysiology signals processing. This project presents two speedy intelligent detection methods of ventricular fibrillation and ventricular tachycardia arrhythmias for ICD devices. It also provides another quick, innovative diagnosis method for use in intelligent electrocardiograph devices to detect abnormal cardiac signals. The first network is 1D-CNN for smart electrocardiographic devices to detect abnormal ECG signals. Dataset MIT-BIH has been used to train this network. This network with the most optimal number of parameters due to high detection speed has a high accuracy of 91%. The second and third networks are 2D-CNNs for use in implantable defibrillators. For the second network, a data set of 20 patients with cardiac arrhythmia and 20 patients without cardiac arrhythmia in an 8-month period of ICD check-up has been prepared. The third network is trained using the Spontaneous Ventricular Tachyarrhythmia Database. The second and third networks are designed to detect EGM signals in VF and VT modes with the optimal number of parameters and 100% accuracy in the second network and 90% in the third network. All three designed networks are in an optimal condition regarding the number of parameters and layers, so they have optimal speed and energy consumption.

Efficient Halftoning via Deep Reinforcement Learning.

Halftoning aims to reproduce a continuous-tone image with pixels whose intensities are constrained to two discrete levels. This technique has been deployed on every printer, and the majority of them adopt fast methods (e.g., ordered dithering, error diffusion) that fail to render structural details, which determine halftone's quality. Other prior methods of pursuing visual pleasure by searching for the optimal halftone solution, on the contrary, suffer from their high computational cost. In this paper, we propose a fast and structure-aware halftoning method via a data-driven approach. Specifically, we formulate halftoning as a reinforcement learning problem, in which each binary pixel's value is regarded as an action chosen by a virtual agent with a shared fully convolutional neural network (CNN) policy. In the offline phase, an effective gradient estimator is utilized to train the agents in producing high-quality halftones in one action step. Then, halftones can be generated online by one fast CNN inference. Besides, we propose a novel anisotropy suppressing loss function, which brings the desirable blue-noise property. Finally, we find that optimizing SSIM could result in holes in flat areas, which can be avoided by weighting the metric with the contone's contrast map. Experiments show that our framework can effectively train a light-weight CNN, which is 15x faster than previous structure-aware methods, to generate blue-noise halftones with satisfactory visual quality. We also present a prototype of deep multitoning to demonstrate the extensibility of our method.

Opt-CoInfer: Optimal collaborative inference across IoT and cloud for fast and accurate CNN inference

For fast and accurate Convolutional Neural Network (CNN) inference of massive Internet of Things (IoT) data, Collaborative Inference (CI) based on partition and compression techniques needs to carefully select the collaboration scheme considering both application scenario and inference requirement. However, with emerging partition and compression techniques, it is hard to select a proper collaboration scheme due to the exponential growth of the scheme space and the time-consuming evaluation of a single scheme. Therefore, we present Opt-CoInfer to search the optimal collaboration scheme for collaborative inference (i.e., the fastest/most accurate scheme satisfying the latency/accuracy requirement in any specific scenario). Generally, Opt-CoInfer achieves the optimum by updating the local optimum and reducing the feasible set to empty in an iterative method. Specifically, in each iteration, Opt-CoInfer selects the promising scheme and updates the local optimum according to a statistical model built with the evaluated schemes, then reduces the feasible set by the local optimum and some key observations of collaboration schemes. Experiments show that Opt-CoInfer outperforms single-end and state-of-the-art CI approaches with the same inference requirements in various application scenarios (e.g., up to 3.49× faster with the same accuracy requirement and as low as 37.41% accuracy loss with the same latency requirement).