China
Africa
Explore chapters and articles related to this topic
Radar Signal Processing: An Example of High Performance Embedded
Published in David R. Martinez, Robert A. Bond, Vai M. Michael, High Performance Embedded Computing Handbook, 2018
A. Bond Robert, I. Reuther Albert
The computational aspects of the SMTI processing stages are discussed next. The subband analysis and synthesis phases, shown in Figure 6-4, are complements of each other. The subband analysis can be implemented by demodulating the incoming signal into a series of subband signals, low-pass filtering each of these subband signals, and downsampling the subsequent filtered signals. The filter inputs are the range vectors of the input radar data cube, as shown in Figure 6-3. Demodulation is performed by mixing or multiplying each sample by a complex downshifting value. The low-pass filtering ensures that signal aliasing is minimized in the downsampling step. Downsampling is conducted by simply extracting every Nth sample. When these operations are completed, the output is a set of subband data cubes, each with the same number of channels and PRIs but 1/Nth the number of range gates. Conversely, subband synthesis entails upsampling the subband data, low-pass filtering the upsampled subbands (using a 15–tap filter in our example), modulating the subband signals back up to their original frequencies, and recombining the subbands by superposition addition. Upsampling is performed by inserting zero-valued samples between each sample. Again, low-pass filtering is used, this time to minimize signal aliasing in the upsampling step. The modulation is performed by multiplying each sample by a frequency upshifting value.
Registration for Super-Resolution: Theory, Algorithms, and Applications in Image and Mobile Video Enhancement
Published in Peyman Milanfar, Super-Resolution Imaging, 2017
Patrick Vandewalle, Luciano Sbaiz, Martin Vetterli
An interesting case is the one where the low-resolution sensor is used to acquire a QCIF video. In this case, the sensor resolution (3.2 Mpixel) is larger than the output resolution (144 × 176 pixels). Blocks of sensor pixels are combined in order to give an effect equivalent to a reduction of resolution (such that the pixel pitch would become 32.7 μm). This operation is equivalent to filtering the image with an averaging filter and then downsampling. Note that in order to avoid aliasing, an (additional) low-pass filter should be applied prior to downsampling. However, handheld devices normally do not include such a filter, and directly subsample the image. It results in a higher level of aliasing on the final images that compose the video. The equivalent response function is shown in Figure 6.2. We remark that a significant amount of aliasing can be present in the range of frequencies between 0.5 and 1 (and even up to 2) cycles/pixel.
Data-driven Detection and Early Prediction of Thermoacoustic Instability in a Multi-nozzle Combustor
Published in Combustion Science and Technology, 2022
Chandrachur Bhattacharya, Jacqueline O’Connor, Asok Ray
The reason why downsampling emerges to be an important hyper-parameter for STSA-based methods is two-fold. Higher downsampling rates lead to poor detection since much of the signal information is not accounted for and is removed. However for STSA, if the signal is not sufficiently downsampled, a situation emerges where the state transition matrices for the STSA method have ‘heavy diagonals’ due to a large number of self-loops within the same state. This is a problem seen when there is oversampling. The fundamental frequencies of the signals are about the 400–700 Hz while the normal signal sampling rate is 16,384 Hz, which even using the maximum downsampling tested (DS = 10) yields a sampling rate of 1640 Hz which is more than two times the highest frequency of interest, thus satisfying the Nyquist criterion. Other than this, there needs to be no mathematical linkage between the fundamental frequencies of the dynamical system and the sampling frequency. In other words, since the best sampling rate in this case is about 8192 Hz (16,384 Hz at DS = 2) when the frequencies of interest are in the range of 400–700 Hz, it means that a desirable sampling frequency would be about 20 times larger than the important frequencies.
Knitted fabric and nonwoven fabric pilling objective evaluation based on SONet
Published in The Journal of The Textile Institute, 2021
Jun Wu, Da Wang, Zhitao Xiao, Kun Yu, Fang Zhang, Lei Geng
The function of the O branch in SONet is to combine the pilling image’s hairball and fabric texture information, which is needed for grading. When using digital image processing for objective grading of fabric pilling, some researchers transform the pilling image into the frequency domain to separate the high frequency component (pilling) from the low frequency component (fabric texture). Inspired by this method, we use an octave convolution to divide the feature maps in the CNN into high and low frequencies (Chen et al., 2019). Because of the spatial redundancy of low frequency components, the memory and computation cost can be reduced by downsampling. Since this branch focuses on replacing the vanilla convolution with an octave convolution, this branch is named O branch.
Multivariate Adaptive Downsampling Algorithm for Industry 4.0 Visual Analytics
Published in Cybernetics and Systems, 2023
Javier Franco, Ander Garcia, Amaia Gil, Juan Luis Ferrando, Xabier Badiola, Mikel Saez de Buruaga
Thus, several downsampling algorithms for data visualization have been proposed (Steinarsson 2013). These downsampling algorithms create a simplified view of the original data, which requires less computation power to be visualized while keeping as much information as possible from the original data. Then, it is possible to visualize data and interact with it, overcoming the technical constraints related to computation or network resources.