Abstract In this paper, we are concerned with image downsampling using subpixel techniques to achieve superior sharpness for small liquid crystal displays (LCDs). Decreasing the number of samples per unit time, sometimes called downsampling, is decimation of the data. Speciﬁcally for ImageNet clas-siﬁcation with the same input size, the proposed method achieves 1.60% and 0.63% top-1 accuracy improvements This implies tremendous savings for coding the difference between the original (unsampled image) and its pre- diction (the upsampled image). The frequency scaling corresponds to having a sampling interval of after downsampling, which corresponds to the interval prior to downsampling. There are two important points to take away about downsampling's effects in the frequency domain: The downsampled signal's frequency spectrum will have its magnitude lowered by the downsampling factor $D$, and will repeat every $2\pi$ Downsampling can cause aliasing. Digital FDM is similar, except the spectrum is repetitive. How to change the sampling rate entirely from the discrete-time domain. Downsampling Section 6, Nick Antipa, 3/9/2018 ... •Compresses in the frequency domain x[n] N y[n] Y (ej! In this paper, we use frequency-domain analysis to explain what happens in subpixel-based downsampling and why it is possible to achieve a higher apparent resolution. \begin{align} \mathcal{X}_2(\omega) &= \sum_{m=-\infty}^\infty s_d[m]x_1[m]e^{-j\omega \frac{m}{D}} \text{ where}\\ &s_d[m] = \left\{ \begin{array}{ll} 1, & \text{for }m\text{ multiples of } D\\ 0, & \text{else} \end{array} \right.\\ &\text{ or }=\frac{1}{D}\sum_{k=0}^{D-1} e^{jk2\pi \frac{m}{D}} \text{ so:}\\ \mathcal{X}_2(\omega) &= \sum_{m=-\infty}^\infty \frac{1}{D}\sum_{k=0}^{D-1} x_1[m]e^{jk2\pi \frac{m}{D}}e^{-j\omega \frac{m}{D}}\\ &= \frac{1}{D}\sum_{k=0}^{D-1}\sum_{m=-\infty}^\infty x_1[m]e^{-jm\frac{\omega -k2\pi}{D}} \\ \end{align}, Comparing to Initially, we have a vector in time domain, consisting of 8 elements, then we transform it in vector of Fourier coefficients, and we are interested in downsampling this vector in frequency domain, such that after the downsampling, we obtain a vector of Fourier coefficients, which has a size 4 in this example. Hence, without using the anti-aliasing lowpass filter, the spectrum would contain the aliasing frequency of 4 kHz – 2.5 kHz = 1.5 kHz introduced by 2.5 kHz, plotted in the second graph in Figure 12-3a. Fig 2a shows data that is nearly oversampled to produce a spectrum that has very little energy in the upper half of the Nyquist interval. The downsampling and upsam- pling schemes combined together preserve all the low- frequency DCT coefficients of the original image. Thus, each of the four frequency bands of Fig 3 could represent separate channels formed by frequency division multiplexing. Let data be an array of 1000 elements. Even so, note that now the Nyquist interval is filled with the nonredundant information that can be used to separate the spectrum of the two channels since and are linearly independent. Figure 12-3A. and wave number k are related by = 2 N.So the spacing in the plot above is! For an oversampling case, the minimum sampling rate is more than 160 MSPS. First, note that when we downsample a signal to a lower sample rate, there is a risk of going below the limit imposed by sampling theorem that can induce aliasing. The idea of downsampling is remove samples from the signal, whilst maintaining its length with respect to time.For example, a time signal of 10 seconds length, with a sample rate of 1024Hz or samples per second will have 10 x 1024 or 10240 samples.This signal may have valid frequency content up to 512Hz or half the sample rate as we discussed above.If it was downsampled to 512Hz then the frequency content would now be reduced to 256Hz, due to the Nyquist theory. frequency domain leverages identical structures of the well-known neural networks, such as ResNet-50, MobileNetV2, and Mask R-CNN, while accepting the frequency-domain information as the input. • Fourier transform. So we cut the high frequency aliases. More later, please send your comments, suggestions, questions etc.