Convolutional neural networks (CNNs) have been widely deployed in computer vision tasks. However, the computation and resource intensive characteristics of CNN bring obstacles to its application on embedded systems. MobileNet, as a representative compact models, can reduce amount parameters computation. A high-performance inference accelerator FPGA for MobileNet is proposed this paper. With respect three types convolution operations, multiple parallel strategies are exploited corresponding...

Identifying the same persons across different views plays an important role in many vision applications. In this paper, we study problem, denoted as Multi-view Multi-Human Association (MvMHA), on multi-view images that are taken by cameras at time. Different from previous works human association two views, paper is focused more general and challenging scenarios of than none these fixed or priorly known. addition, each involved person may be present all only a subset which also not We develop...

The post-training quantization (PTQ) is a common technology to improve the efficiency of embedded neural network accelerators. Existing PTQ schemes for CNN activations usually rely on calibration dataset with good data representation reduce overflow in inference, which not always effective due large variation and uncertainty inference input practice. This paper proposes an adaptive method (AQA), monitors activations, adaptively updates parameters, re-quantizes on-the-fly when degree over...

The post-training compression based on affine quantization is a common technology to improve the efficiency of embedded neural network accelerators. Current state-of-the-art schemes for CNN activations usually rely calibration dataset with better data representation reduce possibility overflow in inference, which not always effective due uncertainty inference input practice. This paper proposes an adaptive method and its hardware-friendly design directly address inference. monitors...

Blockchain is a distributed database with tamper-evident and decentralized data structure. However, due to the huge amount of in blockchain, storage costs are high. In an effort minimize expenses, it necessary choose most pertinent features for preservation on blockchain. Therefore, effective feature selection method can quickly accurately select representative features, which reduce redundancy blockchain upload effectively confirm rights. This paper puts forth algorithm (i.e. FS_NGRE) based...

The work presents an efficient identification threshold signature scheme based on the elliptic curve cryptosystem, with a view to optimizing these techniques for implementation within network. (Any t out of n users in group can represent this sign signature) as compared existing public key system RSA and ELGmal, our need less parameters but provide more super security. Moreover it is constructed two base points other only one point. distinguishing feature proposed that value denotes minimum...

Although statistical modeling techniques have been employed to detect anomaly intrusion and profile user behavior with network traffic data collected from multi-sites (IP addresses), the minimum sample size of audit required for each site is unclear. Using Intrusion Detection Evaluation off-line developed by Lincoln Laboratory at Massachusetts Institute Technology under Defense Advanced Research Projects Agency, this study aimed address challenge determining size. Bivariate analysis was...

Proxy signature can be combined other special to obtain some new types of proxy signature. Threshold is a variant the Most existing threshold schemes based on certificate-based public key systems. But, in cases, receiver cannot find out who signed signatures. In this paper, we put forward novel identity-based scheme from bilinear pairings without dealer. Using our scheme, original signer know generated and certify actuality group signers. Moreover, constructions would simpler but still with...

This paper takes the driver as research object, establishes a vehicle driving data set based on driver's emotions, builds support vector machine (SVM) model to recognize anger.And effects of linear kernel function, polynomial gaussian function and sigmoid recognition result SVM are discussed separately.The results show that is completely unsuitable for anger recognition.The has best effect, which accuracy rate, recall rate F1 score significantly higher than other functions, not much...

In the real world applications, those basic decision rules inducted using rough set theory can't match a new object, so it always make an accurate for object. order to improve matching ability of algorithm which generates approximate more efficiently is proposed in this article and applied field filtering spam based on theory. The results experimental test showed that obtained by can enhance mail. So paper believed be effective feasible.

Abstract To describe rough differential equations in function model, a series of new concepts are proposed by generalizing theory difference and ordinary into model. Based on these basic concepts, fundamental properties linear solutions discussed four principles superposition put forward, which lay foundations for solving methods equations. Several insufficiencies original respectively pointed out improved. Three types typical given. According to characteristics different equations,...

The auto-commissioning system of engineering equipment requires real-time identification the working mode with whole product. Determine whether there is any abnormality in commissioning process according to parameter standards each mode, evaluate compliance rate commissioning. In this article, determined by Toeplitz Inverse Covariance-based Clustering (TICC) and time series data are segmented clustered. Each TICC defined a Markov Random Field (MRF), which characterizes interdependence...

The post-training compression with quantization is a common technology to improve the efficiency of embedded neural network accelerators. In this paper, Dynamic Quantization in Inference (DQI) method proposed solve severe overflow problem that may occur CNN inference process. Based on analysis errors activation values convolutional layers, efficient detection and parameters dynamic update are designed implemented accelerator. evaluation result VGG16 MobileNetV2 models demonstrates DQI can...