|
Compression set sampling perception theory emerged in recent years, compressed and encoded along with the new signal processing method. It breaks through the precise reconstruction of the Nyquist sampling theorem, the use of a very small amount of sample value by solving convex optimal solution, a high probability of the original signal. Due to its characteristics, is now widely used in many practical applications, voice, images, etc.. This paper studies compressed sensing reconstruction algorithm and its application in the neural network vector quantization. The paper first describes the basic theory of compressed sensing, through numerical simulation method to analyze several compressed sensing reconstruction algorithms, including the greedy algorithm Orthogonal Matching Pursuit algorithm (OMP) regularized orthogonal matching pursuit algorithm (ROMP ), subspace matching tracking algorithm (SP), convex optimization algorithm gradient projection algorithm (GPSR) hard iterative threshold algorithm (the IHT), smoothed l 0 norm algorithm (SL0) as well as non-convex algorithm iterative re-weighted algorithm (IRLS), compare the performance of various algorithms. The results show that: the the SL0 algorithm and the IRLS algorithm in reconstruction accuracy was significantly better than the other algorithms, followed by the SP algorithm again ROMP and OMP last GPSR and IHT algorithm. When the sampling rate of 0.9, the standard Lena image, for example, the the SL0 algorithm and IRLS algorithm both peak signal-to-noise ratio (PSNR) of reconstructed images greater than 45dB, the SP algorithm close to the 40dB; purpose of the reconfiguration time, IHT algorithm optimal, under the same conditions, it requires the shortest time, significantly lower than other compressed sensing reconstruction algorithm. Then the paper examines the neural network vector quantization compressed sensing. The vector quantization is an efficient lossy compression technique, when the compression ratio, the image quality will be unsatisfactory recovery. By compression perception theory is applied to the neural network vector quantization, can significantly improve the in compression than large image restoration quality. Paper compressed sensing theory applied to the neural network vector quantization, wavelet layers and measuring the number of different compressed sensing reconstruction algorithms, such as OMP, GP, SP, IRLS, SL0, GPSR vector quantization of the neural network image restoration quality and the impact of the recovery time, by means of numerical simulation, based on compressed sensing neural network vector quantization optimal solution. The results show that: the monolayer wavelet transform using compressed sensing technology enables the lena Figure, girl figure, couple map neural network vector quantization image quality (PSNR) were increased by about 6-7dB ,3-4dB and 1dB below; two layer wavelet transform, improve 3-4dB ,1-2dB and 1dB; under the same conditions, GPSR algorithm has quality and other algorithms used in less time with other algorithms.
|