As an emerging in-memory element, memristor has been widely used in various neural network circuits to represent the weights and accelerate the calculation. However, the Transformer Network (TN), one of the most impor...
详细信息
As an emerging in-memory element, memristor has been widely used in various neural network circuits to represent the weights and accelerate the calculation. However, the Transformer Network (TN), one of the most important models for machine vision and natural language processing in recent years, has not yet been full-circuit implemented using memristors due to the complex calculation process and data storage. In order to carry out the computation of the TN more efficiently, this work proposes a memristor-based full-circuit implementation of the TN capable of: 1) a memristor crossbar module to preserve the weights of the TN and perform the vector-matrix multiplications;2) an analog signal memory module to store the analog signal directly in near-memory mode;3) function circuit modules to achieve five transformations, namely Softmax, Layer Normalization, ReLU, Multiply-add and Residual;4) a timing signal generation module to schedule operations of the circuit. The proposed TN circuit can complete all calculations directly based on the analog signal without using any analog-digital converter (ADC), digital-analog converter (DAC) and digital memory. In addition, character image recognition experiments are carried out in PSPICE to verify the functional correctness of the designed circuit. The corresponding signal retention rates of the analog memory, the performance of the whole circuit, and the non-idealities of the memristors are also analyzed. The results indicate that the circuit has advantages in terms of area overhead, energy efficiency and anti-noise.
The process of medical images using Magnetic resonance imaging (MRI) is important in medical areas;such as the anatomy of human soft tissue, and diagnosis of stroke or cancer. Using mathematical and computational tran...
详细信息
In order to solve parameter estimation problem of Direct spread signal (DS signal) pseudo-code rate in low SNR, the generation mechanism of DS signal and autocorrelation characteristics are deeply studied. The receive...
详细信息
ISBN:
(纸本)9781467395878
In order to solve parameter estimation problem of Direct spread signal (DS signal) pseudo-code rate in low SNR, the generation mechanism of DS signal and autocorrelation characteristics are deeply studied. The received DS signal is processed by follows, such as delay-and-multiply, autocorrelation, low-pass filtering and fast fourier transform, which reduce noise interference. Finally the effective estimation of pseudo-code rate can be realized by searching for the pseudo-code rate spectral lines. Finally by searching for pseudo-code rate spectral lines method can realize the effective estimation of pseudo-code rate. In the case of SNR is -16dB, Simulation results show that the method can realize effective estimation of the pseudo-code rate parameters. And effective estimation of pseudo-code rate in lower SNR can be realized by increasing the processing data points.
作者：
Sun, TaoZhou, LiUniv Jinan
Shandong Prov Key Lab Network Based Intelligent C Jinan Shandong Peoples R China
Depth map prediction is the key point in Free Viewpoint Video technique. Among current depth map estimation approaches, combined temporal and interview prediction method is the most practical one for hardware design a...
详细信息
ISBN:
(纸本)9780769548111;9781467326469
Depth map prediction is the key point in Free Viewpoint Video technique. Among current depth map estimation approaches, combined temporal and interview prediction method is the most practical one for hardware design and real time processing. It is based on block search inter-prediction on various block size to get the best estimation result with specific cost function. Although much data processing bandwidth can be saved by reusing of hardware and computation resources in disparity vector prediction, it still needs to calculate all block size cost results, and has disparity prediction errors at object boundary or continuous areas, resulting in block effect and prediction noises in depth map. This paper presents an efficient adaptive soft decision method based on chrominance image segmentation. The best prediction block size is pre-determined before block search progress. So much calculation efforts are saved. Only specific block size computation is executed to get the best disparity vector prediction, instead of selecting the best one from all block size calculation results. Experiment results show that the adaptive soft decision method can enhance depth map quality efficiently with less prediction errors and computation cost. It is suitable for hardware realization and real time processing.
This paper proposes a noise reduction method of blind source separation based on image sequence. In the image, the signal and the noises can be considered as the independent components, so the multi-frame images can b...
详细信息
ISBN:
(纸本)9781538685273
This paper proposes a noise reduction method of blind source separation based on image sequence. In the image, the signal and the noises can be considered as the independent components, so the multi-frame images can be considered as the multiple linear combinations of one signal and a lot of noises for the noise's randomness. Due to that, the noises can be removed from the sampled multi-images based on blind source separation (BSS). During the separation calculation, the common Gaussian noise is taken as the object to be eliminated, and the nonlinear principal component analysis (NLPCA) is used as the BSS method and the analysis of the noise reduction is based on changing either the noise degree or the image sampling numbers. This proposed algorithm is compared with the multi-frame average (MFA) that is a famous noise reduction algorithm based on image sequence. The analysis results show that the noise reduction properties by using the proposed method will be affected by the noise degree or the image sampling numbers when eliminating the Gaussian noise;and this method can recover the images which are heavily noise polluted, and the suppression effect on strong noises is obviously better than MFA algorithm.
Fuzzy techniques can be applied in several domains of imageprocessing. In this paper we will give a survey on how fuzzy similarity measures can be used in establishing measures for image comparison. Objective quality...
详细信息
ISBN:
(纸本)0780390253
Fuzzy techniques can be applied in several domains of imageprocessing. In this paper we will give a survey on how fuzzy similarity measures can be used in establishing measures for image comparison. Objective quality measures or measures of comparison are of great importance in the field of imageprocessing. These measures serve as a tool to evaluate and to compare different algorithms designed to solve particular problems, such as noise reduction, deblurring, compression, ... Consequently these measures serve as a basis on which one algorithm is preferred to another. Furthermore, it is well-known that classical quality measures, such as the RMSE (Root Mean Square Error) or the PSNR (Peak signal to Noise Ratio), do not always correspond to visual observations.
Unmanned Aerial Vehicles (UAV) digital images are often badly degraded by noise during dynamic acquisition and transmission process. Denoising is very important and difficult for UAV-vision Guided, because natural sce...
详细信息
ISBN:
(纸本)9780769550794
Unmanned Aerial Vehicles (UAV) digital images are often badly degraded by noise during dynamic acquisition and transmission process. Denoising is very important and difficult for UAV-vision Guided, because natural scene image is complicated and having lots of the edges and texture details. The image denoising and enhancement algorithm based on adaptive dual-tree discrete wavelet packets(ADDWP) which combine the dual-tree discrete wavelet packets and evolutionary programming is proposed in this paper. In order to reduce the noise and enhance detail for a UAV image by the EP(evolutionary programming) in wavelet domain(.) Firstly, The de-noising threshold is estimated by EP in the wavelet domain. Secondly, Noise is reduced in the fine high-frequency sub-bands of each decomposition level, respectively, so that the maximum signal-noise ratio can be obtained in the high-frequency subbands, respectively. Thirdly, The enhancement parameters is estimated by EP in the wavelet domain, and the detail is adaptive enhanced by an improved non-linear gain operator in the coarse high-frequency subbands of each decomposition level, respectively. The experiment result improves our algorithm is efficient.
In this paper, a real-time hand gesture recognition system implemented on a TMS320DM642 Digital signal Processor (DSP) of Texas Instruments (TI) is proposed. The highly effective system achieved by combining specialis...
详细信息
ISBN:
(纸本)9781467356046
In this paper, a real-time hand gesture recognition system implemented on a TMS320DM642 Digital signal Processor (DSP) of Texas Instruments (TI) is proposed. The highly effective system achieved by combining specialist algorithms and advanced DSP optimization techniques. The proposed algorithms are composed of face detection and hand gesture recognition together with skin segmentation and hand tracking procedures. Cascade classifiers using AdaBoost algorithm are applied in face detection and hand recognition to accelerate the entire system. In addition, many DSP optimization levels such as floating-point numbers to fixed-point numbers conversion, software pipelining etc. are also exploited. The results show that the proposed design achieves up to 50 640x480 16-bit YCbCr frames per second (fps), which is capable of many real-time applications.
In this paper, it is presented that a new color image watermarking algorithm based on compressed sensing theory and chaos theory in discrete cosine transform (DCT) and singular value decomposition (SVD) domain. Initia...
详细信息
ISBN:
(纸本)9781479970056
In this paper, it is presented that a new color image watermarking algorithm based on compressed sensing theory and chaos theory in discrete cosine transform (DCT) and singular value decomposition (SVD) domain. Initially, a color watermarking image is preprocessed by four steps i.e. being divided into red, green and blue channels, being sparse by wavelet transform, scrambling by 1-D chaos sequence and measurement through Gaussian random matrix. Here, the chaos sequence is produced by 1-D Logistic mapping. Then the red, green and blue channel images of a color host image are divided into 8x8 blocks and then transformed by DCT. Next the first DCT coefficient of each block is selected to compose a new matrix, and the new matrix is decomposed by SVD. The singular values are the regions of watermarking embedment. Thirdly, three sets of measured coefficients as new watermarking are embedded into the corresponding singular values. Finally, the Iterative Hard Threshold (IHT) algorithm is used to reconstruct the watermarking image. The method proposed can not only enlarge the information capacity of watermarking embedment but also strengthen the security of watermarking in the condition of the good imperceptibility. The experiments show that the scheme is feasible and has good anti-attack ability for JPEG compression, Gaussian white noise, rotating and median filtering.
Post-processing is an important module in motion compensated frame rate up-conversion design, and has a direct impact on the image quality of interpolated frame. However, how to balance between image quality and compu...
详细信息
ISBN:
(纸本)9781728103976
Post-processing is an important module in motion compensated frame rate up-conversion design, and has a direct impact on the image quality of interpolated frame. However, how to balance between image quality and computational efficiency is still very challenging for post-processing, especially in hardware design. This paper proposes a hardware-efficient post-processing algorithm which leverages the temporal and spatial constraints to locally refine interpolated pixels. Moreover, we employ the quantization and approximation techniques to further reduce the computational intensity of the proposed post-processing algorithm. The quality of interpolated frame has been greatly improved both in objective and subjective aspects. The proposed post-processing algorithm is extensively evaluated by a set of video test sequences. Evaluation results demonstrate that, compared with the reference designs, the proposed algorithm can improve PSNR by at least 1.69 dB with comparable computational complexity.
暂无评论