No abstract

Radar images obtained by remote sensing systems as a rule are corrupted by mixed additive and multiplicative noise [1]. To carry out high-efficiency filtering, the information about noise statistics in these images is required. However, in many practical situations such information is a priori unknown. Therefore methods able to evaluate variances simultaneously for both noise components are necessary.One of such methods has been proposed in [1]. Its main stages are: image segmentation, homogeneous regions detection, local variance and squared mean estimation for each homogenous region, forming a scatter plot using the obtained estimation pairs, regression line fitting. Slope of fitted line corresponds to multiplicative noise relative variance estimate, and its y-intercept constitutes additive noise variance estimate. General accuracy of the method (bias and estimation variance) substantially depends on the line fitting stage accuracy.That's why in this paper different regression line fitting algorithms are studied:1) fitting along all scatter-plot points using conventional least mean squares algorithm (LMS) [2] and its robust version (RLMS) [3];2) first, centers of scatter-plot "clouds" are determined using robust mode estimation and then regression line is fitted using the obtained reference points by conventional LMS [1] and by weighted LMS (WLMS), where weights are calculated proportional to scatter-plot "clouds" sizes [4].Numerical simulations for a set of test images corrupted by mixed Gaussian additive and multiplicative noise have been carried out. It has been shown that fitting using all scatter-plot points by conventional LMS provides the best accuracy for artificially generated test image, but fails for real-life images because of high sensitivity to essentially biased local variance and mean estimates obtained in heterogeneous image regions.RLMS in majority of situations provides the most unbiased variance estimation for additive component, but its variance occurs to be higher in comparison to other studied algorithms. As for the relative variance estimation, RLMS for low-textured images produces results comparable to conventional LMS, but for highly-textured images its bias is up to several times smaller. WLMS for low-textured images provides biases comparable to the RLMS and sometimes even outperforms it. For highly-textured images, estimates obtained by WLMS are slightly more biased than those obtained by RLMS, but in both cases estimation variance of estimates obtained using WLMS is considerably lower. Hence, WLMS appears to be the most versatile among the considered methods since it is able to provide acceptable evaluation accuracy for images with different characteristics.The considered scatter-plot method is applicable both for pure additive and pure multiplicative noise. In such cases one of the estimates almost equals to zero. The method is also suitable for any complex signal dependent noise, but in this case a polynomial curve should be fitted. Such noise type estimation problem will be stud...

No abstract

This article is devoted to the analysis and processing of signals of inertial measuring modules used as part of inertial navigation systems or in research measuring complexes for conducting shock tests. The modules have a set of sensors for measuring the speed of the object, its orientation in space, the gravitational forces with which it moves, as well as the magnetic field surrounding it. Using the example of a typical inertial measurement module WT901SDCL, it is shown that the signals of acceleration, angular velocity, and angular position generated by the module have a certain fluctuation component, which deteriorates the accuracy of determining the estimated parameters, as well as to the appearance of an accumulated error in the determination of coordinates of the inertial navigation system. In such conditions, it is advisable to use secondary processing methods, namely, filtering methods. Since the information component of the signals of the inertial measurement module has significant dynamics, one of the key requirements for the filtering method is its preservation. However, a filter is also required to effectively suppress the fluctuation component. Among the currently existing filters, the filter based on discrete cosine transform (DCT-filter) has the best trade-off according to these requirements. It is shown that the use of this filter allows reducing the intensity of the noise component in the information signal by the average of 1.4 times the value of the mean square error and, accordingly, to reduce the measurement errors of physical parameters. Simultaneously, the shape of the signals after applying the DCT-filter remains almost unchanged, all sharp dynamic changes in the signal are preserved, and the absolute levels of the signals also remain the same. Thus, the use of DCT-filtering for signal post-processing in inertial measurement modules can be considered quite reasonable.